H3lite

MiniMax H3 Local Video

Use this skill to turn a user's local computer into a reproducible MiniMax H3 audio-video workstation and to generate short clips. The validated primary route is Windows + NVIDIA + ComfyUI; macOS Apple Silicon is a community/experimental alternative, not an equivalent tested backend. Treat platform selection, path selection, hardware-aware planning, prompt writing, execution, timing, and verification as one workflow. The default is the validated fast route; other modes must be chosen explicitly or justified by a time budget.

Platform scope and routing

Detect the operating system and accelerator before giving installation commands or model links. Do not give CUDA, Windows virtual-environment, .bat, or Windows path instructions to a macOS user.

For a non-Windows user, preserve the useful cross-platform guidance (prompt structure, 32-pixel canvas alignment, low-resolution preview, disk/log/output checks), but label every resource and timing as platform-specific. A community implementation that can run H3 on a Mac is evidence for feasibility, not evidence that this ComfyUI skill supports Metal.

The Mac alternative described in the project notes uses MLX/Metal and a third-party mmh3turbo package. If a Mac user explicitly chooses that route, point them to the author's community bundle and package (uvx mmh3turbo), and state that these are external resources with independent licensing, updates, and validation. It may reduce the download footprint with GGUF/4-bit bundles, but it is outside this skill's tested component sets. Do not silently install it, mix its weights with ComfyUI models, or present its 30–43 minute 5-second 720p timings as Windows benchmarks.

Agent production workflow

For complex creative requests, use the compact workflow contract in references/agent-workflow.md:

intent route → reference/identity anchors → prompt enhancement → execute → verify

Route from the user's input and acceptance criteria, not from a style adjective. Use I2VA for a specified opening frame, FL2VA/L2VA for endpoint anchors, and Ref2VA only after its model, text encoder, and workflow are confirmed. For recurring characters or multi-shot work, define stable subject/reference labels, what must be retained, what may change, and what drift is forbidden before writing the timeline. This workflow pattern is local and does not add a cloud service, MCP dependency, second model, or second inference pass.

When a creative brief is vague (for example, it only says “more cinematic” or “make it a nice 3D animation”), optionally read references/prompt-assist.md. It adapts the public Higgsfield prompt structure—stable style/identity lock, one clear scene action, physical camera motion, audio, and compact anti-drift constraints—to H3's native fields. Use it as a writing aid only: do not call Higgsfield, copy model-specific flags or capabilities, or let a web lookup change the local route, resolution, component set, or verification rules. If a live lookup is needed, follow the host web-access skill and use public pages only.

Operating rules

• Hot path first: for an ordinary text-to-video request on an already validated installation, run scripts/h3_fastpath.py once. It combines /system_stats, fresh-cache reuse, in-process planning/preflight, queue submission, and one bounded completion watch. Do not issue repeated one-shot status calls, reread the full reference set, run --help, or ask nonessential questions during this path. If the command yields a running terminal cell, wait on that cell; do not start another monitor.
• Keep cold work out of the hot path: model download manifests, hash or size verification, Torch/custom-node repair, repository checks, browser workflow discovery, and full recursive doctor scans belong to installation, migration, repair, or first-run validation. A normal generation on an unchanged machine must not pay those costs.
• Cold path can be heavier when it prevents hour-scale waste: during installation or repair, verify download source, target folder, expected size/hash when available, runtime imports, and model-role mapping before queueing. Store the result so later prompts reuse it instead of repeating it.
• Inspect before changing anything. Run scripts/h3_doctor.py --json and locate the target ComfyUI directory before installing packages, nodes, or weights.
• On first contact, run a platform/accelerator check before the Windows doctor. If the machine is not Windows + NVIDIA, stop the CUDA installation branch and route the user using the platform matrix above.
• Prefer an isolated ComfyUI directory when no installation is supplied. Never overwrite an existing installation or silently replace model files.
• Keep the deployment path configurable. Do not copy paths from another computer into scripts or workflows.
• Report required disk space before large downloads. Use resumable downloads and verify file size or hash when a source provides one.
• Before a long generation or multi-shot batch, check free space and pagefile headroom and keep per-shot logs. A pipeline that filters away the process exit code or traceback is not a successful run; preserve the full log and stop on the first failed shot.
• If the user can only download from the public internet, run a cold-path download plan before fetching multi-GB files: test candidate raw URLs with a small ranged download, choose the fastest stable source, estimate wall-clock time, then use resumable .part downloads. Do not pretend scripts can beat the user's real bandwidth.
• Before downloading large assets, run the doctor compatibility probe. Stop on a Torch import error; treat a comfy-kitchen/Torch mismatch as a repair decision, not a post-download surprise. Do not silently substitute model files or start unlimited parallel downloads.
• Treat the diffusion checkpoint, text encoder, Turbo LoRA, workflow, and node revisions as one component set. Read references/component-sets.md during installation, migration, model replacement, or kernel repair. Never construct an unvalidated set from individually plausible filenames.
• Use --component-set auto for one unambiguous installed set, or explicitly select A/validated-low-vram-a or B/portable-16gb-b when both sets are installed. Record the selected set in the run manifest; never resolve a partial set role by role.
• Prefer the maintained Baidu package for the registered A/B component sets. Keep the selected set atomic, and respect the licenses of model weights and third-party nodes when using either the package or upstream sources.
• Prefer the ComfyUI HTTP API with an API-format workflow JSON. Use browser/CDP capture only as a recovery path when no reusable workflow JSON exists.
• Check http://127.0.0.1:8188/system_stats before starting anything. If ComfyUI is already healthy, reuse it and do not restart it or rediscover its workflow history.
• Preserve MiniMax H3's audio path and flow/sigma-shift handling when the user wants native audio. Do not remove audio VAE, audio conditioning, or the H3 sampling node merely to make a graph look simpler.
• Zero-inference optimization constraint: hardware compatibility checks, timing calibration, face routing, and media QA may run before or after generation, but must not add sampling steps, extra generation models, or a second video inference pass. Keep the selected graph unchanged unless the user explicitly requests a different quality profile.
• Face-quality routing: if the user needs a recognizable or speaking human face, do not treat low-VRAM W4A8 T2VA at 640x352 as a final-quality route. Prefer I2VA with a clear first-frame reference; prefer Ref2VA when identity must persist across shots. Read references/face-quality.md, confirm MiniMaxH3ReferenceToVideo through /object_info, and confirm the matching reference-capable text encoder/projection and workflow before selecting that route. A registered node alone is not enough; the bundled Ref2VA templates are an experimental local path until a complete run passes media and manual identity QA.
• Anchor before prompt: for multi-shot or identity-sensitive requests, first create an internal anchor sheet with stable subject/picture labels, retention rules, allowed changes, and forbidden drift. Use the same labels in the prompt, output prefix, and run manifest; read references/agent-workflow.md for the compact contract.
• Assist vague creative briefs without inventing facts: when the request lacks a concrete camera, action, sound, or finish, read references/prompt-assist.md and use its bounded defaults or ask one targeted question if the omission changes the route or acceptance criteria. A public Higgsfield lookup is optional and pattern-only; fall back to the local references when browsing is unavailable.
• On current ComfyUI builds, the API class MiniMaxH3SigmaShift is the native ModelSamplingMiniMaxH3 node and uses the merged ModelSamplingAV video/audio schedule fix. Detect it by /object_info or the local source before adding a custom dual-clock sampler; do not duplicate the fix merely because the API class keeps its compatibility name.
• Run the read-only planner before a non-trivial generation. It must report selected mode, resolution, steps, cache policy, paths, and an estimated time range. Do not present an estimate as a guarantee.
• Run the read-only preflight after the doctor and planner. Treat low available RAM/VRAM as a caution, but stop when the pagefile is critically low, required assets are missing, or the doctor recommends an alternative backend.
• Do not perform a full recursive doctor scan for every prompt. Cache the environment report under <ComfyUI>/user/h3lite_runs/_environment/; reuse it for a normal session (normally no older than 30 minutes), invalidate it after ComfyUI/model/node/driver changes or a failed run, and use h3_preflight.py --refresh-runtime for volatile resource fields.
• Do not revalidate large model files before every prompt. Trust the cached download/component manifest unless the file is missing, has a different size/mtime than recorded, the user changed components, or the previous run failed with a model/node/loader error.
• For registered Set B files, require the recorded SHA-256 on first use or after a size/mtime change. A same-size corrupted W4A8 checkpoint produced colored mosaic frames, so byte count alone is not proof of integrity; reuse the cached integrity result on unchanged files.
• Treat every submission as an auditable run: save the effective prompt, mutated API workflow, configuration fingerprint, queue ID, actual execution time, and verified output in the run manifest.
• For identity-sensitive or multi-shot runs, the runtime also writes anchors.json beside manifest.json and records advisory anchor_qa comparisons; these signals support manual continuity review but are not face recognition.
• Keep agent-facing status compact: omit ComfyUI's full history graph by default; use verbose history only when diagnosing a failure.
• Never submit an identical configuration while its manifest is submitting, queued, or running. Return the existing prompt ID instead; use --allow-duplicate only when the user explicitly asks for a second identical run.
• Treat low-VRAM timing as an empirical estimate. The first run can be much slower because kernels compile and weights move between system RAM and VRAM.
• For expensive renders, use a cheap preview pass first: validate the complete prompt/shot list at the smallest supported canvas (for example 256p or the local fast bucket), then promote only approved shots to the requested resolution. This is especially important for multi-shot work; it is a planning optimization, not a second quality-generation pass for a single requested clip.
• Prefer NORMAL_VRAM when a validated 16 GB system can keep Set B resident. In a same-model/workflow/prompt/seed 640x352 comparison, an RTX 4060 Ti 16 GB run took 77.08 seconds versus 591.22 seconds on an RTX 4070 Laptop 8 GB using LOW_VRAM; treat dynamic loading/offload as the main operational explanation, not as a pure GPU benchmark or a promise.
• When launching ComfyUI as a background process, redirect stdout and stderr to persistent files. A detached pipe can become invalid after the launching session is cleaned up, leaving ComfyUI alive but causing tqdm/logger writes to fail with OSError: [Errno 22] Invalid argument. On that signature, restart ComfyUI with persistent logs; do not redownload models or rerun a full doctor unless the restart exposes another error.
• Keep media verification attached to the selected ComfyUI root. The verifier searches system PATH, H3LITE_FFPROBE, and common locations in or beside <ComfyUI> for ffprobe; both h3_generate --watch and standalone h3_status must receive or infer that root. Standalone status may infer the parent only when --output-dir points exactly <ComfyUI>\output; otherwise pass --comfyui explicitly. Treat ffprobe_not_found as a missing verifier, not evidence that generation failed, and do not requeue the video until the existing output has been inspected.
• Treat run-history cleanup as explicit maintenance, never hot-path work. Use scripts/h3_cleanup.py in dry-run mode first and require --apply before deleting eligible run snapshots. Preserve _environment, _hotpath, _workflows, _experiments, prompt folders, timing data, and generated output files.

Preferred component download source

For installation or repair, use the maintained Baidu package before assembling the set from multiple upstream repositories. Select one complete package after the hardware check; do not ask the user to download both sets or mix their exclusive files.

Guide the user to open the matching link, enter the code, and download the whole package. If the baidu-drive skill or a Baidu Drive connector is available, use it for the download; otherwise give the link and code directly and continue after the user places the files locally. Merge the package's models and custom_nodes folders into the selected <ComfyUI> root, then import or copy the packaged workflows and keep component-manifest.json with the install. Run the doctor after the merge.

Set A contains the INT4 text encoder and optional low-VRAM acceleration nodes. Set B contains the FP8 text encoder and validated compatibility workflows. Both packages include their own shared ClipProj and VAE files, so a user only needs one link. If the Baidu package is unavailable or the user explicitly requests upstream downloads, use the exact sources, filenames, sizes, and hashes in references/component-sets.md.

Installation target contract

Before installing, downloading, or moving any component, establish one explicit deployment target and state it to the user:

Install mode: reuse-existing | current-project | dedicated-folder
ComfyUI: <absolute path>
Models: <ComfyUI>\models
Custom nodes: <ComfyUI>\custom_nodes
Output: <ComfyUI>\output

Use these rules:

• reuse-existing: use the exact existing ComfyUI path supplied by the user or discovered and confirmed by the user. Do not clone, reinstall, or create a second model directory.
• current-project: keep everything under the active workspace in <workspace>\.h3lite\ComfyUI so the project-scoped choice is unambiguous and does not scatter models across the repository.
• dedicated-folder: use the user's absolute path, preferably a non-repository path such as D:\AI\MiniMax-H3\ComfyUI or F:\MiniMax-H3\ComfyUI. Put the venv, custom nodes, models, user data, and output under this ComfyUI root.
• If no existing installation and no target path are available, recommend dedicated-folder and ask the user to confirm the absolute path before downloading large files. Never silently choose a drive or install into the current project root.
• If the user says “当前项目” without naming the workspace, resolve and display the active workspace path before proceeding. If the user gives a path ending in ComfyUI, use it directly; if they give a parent install folder, append ComfyUI and display the resulting path for confirmation.

After the target is selected, read references/deployment-matrix.md and present the component checklist, exact destination folders, estimated disk budget, and the launch command. Install or repair in this order: runtime prerequisites, ComfyUI, required custom nodes, model files, doctor verification, then launch. Do not start a generation while any required node class or model asset is missing.

Adaptive planning contract

Before generation, separate the user's requirements into four independent choices:

• Intent: fast (default), balanced, or quality.
• Wall-clock budget: maximum expected generation time, such as “within 10 minutes”; this is different from the requested clip duration.
• Canvas: auto, landscape/16:9, portrait/9:16, square/1:1, or an explicit WIDTHxHEIGHT. Keep pixel size as a ComfyUI parameter, not as prompt prose. The prompt should describe framing such as landscape composition, medium close-up, macro shot, or slow push-in.
• Target path: reuse-existing, current-project, or dedicated-folder.

Run the planner after the doctor scan. It is read-only and does not create folders:

python scripts/h3_plan.py `
  --root <hardware-or-disk-root> `
  --comfyui <ComfyUI-path> `
  --install-mode reuse-existing `
  --mode auto `
  --target-minutes 10 `
  --aspect landscape `
  --megapixels 0.4 `
  --video-seconds 5 `
  --json

Use the planner result as the source of truth for --profile, --resolution, --steps, --length, and --fps. If no quality or time requirement is given, auto selects fast and preserves the validated 640x352, 4-step, Block-Cache route. On an 8 GB laptop, balanced keeps 640x352 and increases steps while bypassing T8 Block Cache; quality does the same with 8 steps. Only the mid/high-VRAM plan promotes balanced/quality to an official-style ComfyUI bucket by default.

ComfyUI's native H3 templates commonly use ResolutionSelector: aspect ratio

• target megapixels + a 32-pixel multiple. For example, 16:9, 0.4 MP, multiple=32 gives 864x480. Treat that as the normal official-style H3 canvas, while H3 Lite's 640x352 remains the low-VRAM fast baseline. If the user asks for "official template size", "normal quality", or "0.4MP", use --megapixels 0.4 unless hardware/time preflight blocks it.

The 32-pixel alignment is a practical model/decoder constraint: the VAE's 16-pixel spatial reduction and the DiT's 2-pixel patching must both align. Do not advertise consumer labels such as "720p" as exact model sizes when the nearest legal canvas is different; report the actual canvas (for example 1280x704) and keep the requested label as a human-friendly preset name.

For a prompt with several ordered actions, prioritize adherence before pixels. Use the fast 640x352 baseline first; 736x416 is an experimental adherence bucket between the fast canvas and 0.4 MP. Do not assume 768p follows complex shot structure better: community reports describe stronger Ref2VA/keyframe adherence around 352p-416p, while extra steps mainly improve coherence and detail. Promote 736x416 only after a same-prompt comparison succeeds locally.

When the user supplies an explicit resolution, treat that canvas choice as already confirmed: issue one concise OOM/time warning, then continue unless preflight is blocked. Do not ask the same resolution question again. If a separate wall-clock budget conflicts with the explicit canvas, state the conflict once and keep the user's explicit canvas. Never silently trade away the audio path or change the installation target to make the estimate fit.

Fast path for ordinary text-to-video

Use this path for a short H3 clip on a machine that passes the low-VRAM doctor check. With no frame arguments it is T2VA; adding --first-frame or --last-frame selects the native reference route automatically:

python scripts/h3_fastpath.py `
  --comfyui <ComfyUI-path> `
  --prompt-text "<rewritten H3 prompt>" `
  --resolution 640x352 `
  --video-seconds 5 `
  --filename-prefix video/H3Lite_my_clip `
  --json

For image-to-video, pass the reference image directly. The helper stages an external image into <ComfyUI>/input using a content-addressed filename, then connects it to the native MiniMaxH3ImageToVideo node. No extra I2V node or manual JSON editing is required:

python scripts/h3_fastpath.py `
  --comfyui <ComfyUI-path> `
  --mode i2va `
  --first-frame <path-to-first-frame.png> `
  --prompt-text "<rewritten I2VA prompt>" `
  --resolution 640x352 `
  --video-seconds 5 `
  --filename-prefix video/H3Lite_i2va `
  --json

Use --mode fl2va --first-frame <first> --last-frame <last> for fixed first and last images, or --mode l2va --last-frame <last> for a last-frame route; the helper removes the I2V template's first-frame placeholder automatically in the latter case. When --mode auto is left in place, the mode is inferred from the supplied frame arguments. The low-VRAM baseline remains 640x352, 124 frames, 4 steps, and native H3 audio. A validated RTX 4070 Laptop 8 GB I2VA run at that baseline took about 12 minutes; treat that as local empirical timing, not a guarantee.

已配置环境的复跑路径就是这一条命令；不要再分别调用 doctor、plan、preflight、generate 和 status。Windows/Git Bash 路径请写 F:/MiniMax-H3/ComfyUI，不要写 /f/MiniMax-H3/ComfyUI。

The helper reuses <ComfyUI>/user/h3lite_runs/_environment/doctor.json for 30 minutes, invalidating it only for a cold start, explicit --force-doctor, an installation change, or failure recovery. It writes the plan and prompt into the run root, keeps the native audio path unless the prompt explicitly asks for complete silence, and waits with one compact --watch monitor. An explicit resolution is already confirmed; continue after one concise risk warning unless preflight is blocked. --dynamic-check is on by default; use --skip-dynamic-check only for an intentionally static clip.

Hot-path budget rule: spend only the time needed to rewrite the prompt, refresh volatile runtime status, submit, watch, and verify. Do not check Git remotes, download pages, model hashes, dependency versions, or official docs during a normal generation unless the previous command returns a concrete error pointing there. If cache is valid, proceed directly to generation.

Native Windows progress window

On Windows, the fastpath opens the native monitor by default, so every normal desktop generation has a visible progress window without an extra flag. Use --no-monitor-gui for a run that must stay terminal-only; --monitor-gui explicitly forces it on. The window discovers the fresh H3 run manifest automatically. It uses a native Windows Tkinter window, reuses the manifest's ComfyUI client_id, and listens to the native /ws channel, including the newer progress_state node events, while HTTP polling supplies queue state, elapsed/estimated time, GPU memory, RAM, pagefile, output path, and failure state. It is monitor-only: closing it does not interrupt generation. The node count is structural workflow progress, not elapsed-time progress: H3 nodes have very different runtimes. The window therefore shows node completion separately, displays the current node's observed runtime, and keeps ETA on the empirical timing estimate instead of treating 4/5 as 80% of the time. The track is segmented by workflow node so completed, active, and pending nodes remain visually distinct. The default window is 760x620; its content area has a vertical scrollbar and mouse-wheel support for smaller displays or larger system scaling. The monitor JSON marks these semantics as progress_basis (node_completion or sampling_steps) and eta_basis (empirical or live_progress) so an Agent can report them without inventing a time percentage. When available, elapsed time comes from the run manifest's measured execution time; the wall-clock timestamp is only the fallback before that field exists. Old running manifests are ignored during automatic discovery; pass --prompt-id to inspect a specific historical run.

To open the monitor independently:

python scripts/h3_monitor_gui.py `
  --comfyui <ComfyUI-path>

Use --once --no-websocket for a one-shot JSON diagnostic. If the optional WebSocket client is unavailable, the window remains usable through HTTP polling, but the progress track stays static and explicitly reports that live quantifiable progress is unavailable. It never substitutes an animated bar for a measured percentage. An MCP wrapper is unnecessary for this local GUI; an Agent can query the same monitor JSON separately when it needs status.

Use the lower-level h3_doctor.py, h3_plan.py, h3_preflight.py, h3_generate.py, and h3_status.py commands only for installation, recovery, custom workflows, or diagnosis. Never replace the fastpath with repeated one-shot status commands during a normal generation.

Use the history/object-info/browser fallback only when the bundled template is incompatible, a Ref2VA/custom graph is needed, or the queue reports a missing node/model. T2VA, I2VA, FL2VA, and L2VA are now first-class bundled routes.

Workflow

1. Diagnose the computer

Run:

$environment = '<ComfyUI>\user\h3lite_runs\_environment'
python scripts/h3_doctor.py `
  --json `
  --report-file "$environment\doctor.json" `
  --root <chosen-root> `
  --comfyui <ComfyUI-path>

Record GPU name and VRAM, system RAM, free disk, Python, CUDA/PyTorch visibility, ComfyUI location, model presence, and custom-node presence. If there is no NVIDIA CUDA device, do not promise this local CUDA route; explain the limitation and offer API/cloud or another backend as an alternative.

The doctor also records available physical RAM, available Windows pagefile, and GPU compute processes. Low available RAM or VRAM is a warning because the validated 8 GB route can still finish with offload. A nearly exhausted pagefile is different: it previously caused hostbuf_file_reader_read failed and system-level paging failures, so h3_preflight.py blocks that run.

Use references/deployment-matrix.md to choose a profile. For an 8 GB laptop, start with the tested W4A8 profile, not the larger official INT8/32B profile.

If the user asked for installation or repair, report the selected installation target before the doctor result. The doctor is read-only; it does not install ComfyUI, nodes, Python packages, or model weights.

2. Select a profile

• Fast (default): a registered W4A8/4B/Turbo component set, 4B ClipProj, FP16 video VAE, FP32 audio VAE, 640x352, 124 frames, and 4 steps. The launch profile uses --lowvram for the very-low/8 GB tiers; 10–16 GB systems use normal VRAM mode unless preflight or a prior OOM justifies offload. Use Block Cache only when its classes are actually loaded; otherwise use the compatibility workflow. This is the success-rate baseline.
• Balanced: keep the low-VRAM canvas on an 8 GB laptop, use 6 steps, and bypass Block Cache. On a mid/high-VRAM machine, the planner may select 864x480.
• Quality: use 8 steps and bypass Block Cache. On an 8 GB laptop, keep 640x352 and warn that W4A8/4B remains a quality ceiling; on a mid/high-VRAM machine, the planner may select 864x480.

The accelerated fast graph is paired with the registered LightX2V/Turbo 4-step LoRA. minimax_h3_turbo_v4_step600_ema.safetensors is a compatibility-workflow quality variant: H3 Lite rejects it when Sage/Sol/Chunk/T8 acceleration nodes are present, because that combination produced severe ghosting and color artifacts in local validation. Use a *_compat_api.json workflow for v4, or use the registered LightX2V/Turbo 4-step LoRA for the fast route.

• 6 GB experimental: when the machine has roughly 6 GB VRAM, 32 GB system RAM, an SSD, and sufficient pagefile headroom, permit a cautious first run at 608x352, 4 steps, and low-VRAM offload. Treat community timings as orientation only: reported I2V runs include about 345 seconds at 608x352/4 seconds and 441 seconds at 864x480/5 seconds, while another 640x480/5-second report took about 13.7 minutes. These used different official/community model and workflow combinations, so do not transfer the numbers to the bundled W4A8 graph as a promise.
• Below roughly 6 GB or insufficient RAM/disk: stop before downloading or queueing. Explain the missing capacity and propose a hosted/API or alternative model.

Do not infer that a smaller file or INT8 label is automatically faster. On low-VRAM systems, CPU offload, RAM bandwidth, kernel compatibility, and first-run compilation often dominate.

The planner's --target-minutes option chooses the highest-quality mode whose conservative upper estimate fits the budget. Without that option, auto always selects fast. The estimate must include a cold-start warning and the final report must include ComfyUI's actual execution time.

3. Install or repair the environment

Use the official ComfyUI H3 tutorial and the component list in references/deployment-matrix.md. Create or reuse the selected ComfyUI root and place every component below that root. Pin or record the ComfyUI and custom-node commits used for a successful run. Install only nodes referenced by the selected workflow. The baseline requires ComfyUI-ClipProj or a compatible implementation. KJNodes, H3 Turbo helper nodes, Sol Attention, and T8 Block Cache are optional unless the selected accelerated graph explicitly uses them.

h3_fastpath.py --workflow-template auto is the default. It queries /object_info when available and uses the accelerated T2V or I2V graph only when the Sage, Sol, Chunk Feed Forward, and T8 classes are actually loaded. Pass --component-set A or --component-set B when both complete sets are installed. Set B is validated with the compatibility graph on RTX 4060 Ti 16 GB and RTX 4070 Laptop 8 GB systems, so auto selects that graph. Its full Sage/Sol/Chunk/T8 acceleration chain is not yet the validated default; use --acceleration fast only for an intentional trial. Use --acceleration compat to force the validated compatibility graph. Both graphs preserve the H3 sampler, native audio, ClipProj, LoRA, dual VAEs, and native first/last-frame inputs without optional patches.

The bundled multi-image Ref2VA graphs reuse this same registered component set: there is no separate Ref2VA checkpoint to download. They add the native MiniMaxH3ReferenceToVideo route and bind repeated reference images through the resident ClipProj path. Before downloading anything, reuse the local W4A8, 4B encoder, ClipProj, dual VAE, and Turbo LoRA files when their manifest and loader checks pass.

Keep optional INT8 loaders and experimental cache nodes disabled until the baseline works. Start ComfyUI with a profile-appropriate command; the very-low/8 GB launch profile commonly uses --lowvram and --fast-disk, while 10–16 GB systems normally omit --lowvram unless preflight or an earlier OOM justifies it. Add Sage Attention only after its PyTorch/CUDA compatibility is confirmed. Treat Easy Cache and generic cache nodes as opt-in experiments: community reports and local experience show that some settings can blur or damage motion/detail. Never enable a cache solely from a speed claim; compare a short output against the uncached baseline first.

After installation, rerun the doctor and stop if any required model or node is missing. Do not start a long generation while the graph contains unresolved node classes.

Cold-path validation is deliberately allowed to spend extra seconds or minutes: record component URLs, selected replacement filenames, file sizes or hashes, local loader choices, Python import probes, and any small compatibility patch that was applied. This is cheaper than discovering a wrong 20 GB model or a Torch/custom-node mismatch after queueing. Once the baseline succeeds, write or refresh the cached environment/component report and return to the hot path for future prompts.

When public internet is the only source for a large model such as the W4A8 diffusion checkpoint, do not start a blind long transfer. First probe the official raw URL, configured mirror raw URL, and any user-provided proxy URL with a small resumable/ranged request. Report the fastest measured speed, estimated time, and selected command. Prefer aria2c with conservative connections when available and the server supports ranges; otherwise use curl.exe -L -C -. Keep partially downloaded files as .part and resume them; do not delete progress or restart from byte zero unless the size/hash proves the file is unrecoverable.

Before queueing, run:

python scripts/h3_plan.py `
  --doctor-json <ComfyUI>\user\h3lite_runs\_environment\doctor.json `
  --install-mode reuse-existing `
  --comfyui <ComfyUI-path> `
  --mode auto `
  --aspect landscape `
  --video-seconds 5 `
  --report-file <ComfyUI>\user\h3lite_runs\_environment\plan.json `
  --json

python scripts/h3_preflight.py `
  --doctor-json <ComfyUI>\user\h3lite_runs\_environment\doctor.json `
  --plan-json <ComfyUI>\user\h3lite_runs\_environment\plan.json `
  --refresh-runtime `
  --require-audio `
  --json

ready means no observed risk, caution means the run may proceed with an explicit warning, and blocked means fix the environment or use another backend first. The preflight ignores the Python process used by ComfyUI and zero-memory desktop helpers; it warns only for meaningful external GPU competitors. This is a runtime gate, not a promise of success.

4. Rewrite the user's request into an H3 prompt

Identify the generation mode before writing:

• text only → T2VA
• one first-frame image → I2VA
• first and last images → FL2VA
• last-frame image → L2VA
• reusable images/video/audio references → Ref2VA

Read references/prompt-writing.md when composing or revising a prompt. For multi-shot, identity-sensitive, or reference-heavy requests, also read references/agent-workflow.md and build the route/anchor sheet before writing. For an underspecified creative request, also read references/prompt-assist.md: it supplies the optional STYLE/IDENTITY LOCK → SCENE → MOTION → AUDIO → NEGATIVE scaffold and the rules for translating it back to H3. Ask only for information that materially changes the route or acceptance criteria; otherwise state the bounded default and continue. Never make a website lookup a runtime dependency. Use the exact field names and ordering required by the selected mode. For native base modes, the core order is:

integrated_multimodal_description: ...
overall_soundscape: ...
non_diegetic_music: ...

For user-facing 5-second quick-start examples, teach the same idea as a memorable three-part structure: scene and atmosphere → action and camera → sound. Present it as one natural-language prompt that can be copied directly; do not require users to write schema labels. Treat this as an explanation aid, then translate it internally into the workflow's required prompt schema.

For full-reference mode, use the six-section structure in the reference. Write the rewritten description in English, but preserve user-provided dialogue, lyrics, and visible scene text in the original language.

Make the prompt operational:

• establish style, framing, subjects, environment, lighting, and initial state in Shot 1;
• make each important person's orientation relative to the camera explicit when identity or facial visibility matters; “watching the sunset” alone often implies a back view. State front-facing, three-quarter, profile, or back-facing, and say whether the face and eyes must remain visible;
• describe observable actions and state changes in playback order;
• use later shot cut times only when a real cut introduces new information;
• write camera motion as a natural sentence with motion type, amplitude, and speed when useful;
• describe dialogue, singing, and diegetic sound in the timeline body;
• describe ambient/physical sound in overall_soundscape and audience-only music in non_diegetic_music;
• Resolve the audio policy before submission: auto/ordinary prompts require the native H3 audio path; “no dialogue” does not disable sound. Only explicit complete-silence wording may remove the CreateVideo audio input.
• Interpret 不要对白 / 无对白 / no dialogue as a dialogue-only constraint: keep native ambience, sound effects, animal or action sounds, and the audio stream. Disable all audio only when the user explicitly requests 完全静音, 无任何声音, or no audio.
• use stable speaker IDs and exact text for dialogue;
• use N/A for music only when no non-diegetic music is desired, and use overall_soundscape: N/A only for explicitly complete silence.

If the selected ComfyUI graph uses a ClipProj/krea2 or another custom prompt schema, inspect its example workflow first. Adapt the official semantic structure to the node's accepted field while preserving the graph's required fields; do not blindly paste a T2VA prompt into a Ref2VA input or vice versa.

• For a face-quality attempt, constrain the first shot to one face, a front or three-quarter orientation, visible eyes, stable hair/clothing anchors, and static or small-amplitude motion. Treat dialogue as a second-stage stressor. A valid MP4 with dynamic/color/audio checks can still have unusable faces; visually inspect face consistency at first/middle/last frames.
• For Chinese prompts, avoid an extremely short noun-only description. H3's long multimodal sequence can make a one- or two-token Chinese prompt easy for the seed to dominate. Add concrete subject traits, setting, framing, lighting, and motion; as a practical starting point, use roughly 30–50 Chinese characters or an equivalent amount of structured detail, then validate with a cheap preview.
• For a director-level multi-shot sequence (establishing + over-the-shoulder + reverse + reveal) where H3's single-continuous-shot limit blocks real coverage, see references/director-sequences.md. It covers splitting the scene into separate I2VA segments, generating consistent first-frame images with ImageGen (watermark crop, reference-image identity inheritance), locking character identity across prompts, W4A8 skin-quality phrasing, and stitching with ffmpeg xfade/acrossfade.

5. Generate through ComfyUI

For installation, recovery, or custom-workflow debugging, the lower-level helpers remain available. A normal text-only request should use h3_fastpath.py above:

python scripts/h3_generate.py `
  --base-u

…(truncated)