Sogni Creative Agent Skill

Sogni Image, Video & Music Generation

Generate images, videos, and music using Sogni AI's decentralized GPU network through the sogni-agent CLI.

Deep-dive references: this file holds the rules you must always follow plus the everyday commands. Detailed guides live in references/ — read the matching file before acting on those tasks (table at the end of this file). If the references/ directory is not present in your install, run sogni-agent --help for the full flag reference or fetch the guides from https://raw.githubusercontent.com/Sogni-AI/sogni-creative-agent-skill/main/references/.

Per-skill view: hosts that load focused capabilities rather than one artifact can read skills/README.md for the per-skill index of the hosted tool surface.

Install Request Policy

When a user asks to install this plugin or skill, install the command-line tool plus this skill:

npm install -g @sogni-ai/sogni-creative-agent-skill@latest
sogni-agent --version

Then configure the agent/runtime to use this SKILL.md and invoke the sogni-agent CLI. The one-command alternative npx setup-sogni-agent-skill auto-detects Claude Code, Codex CLI, and Hermes (it does not configure OpenClaw).

After any install or upgrade, verify with:

sogni-agent doctor

Agents should run sogni-agent doctor --json and confirm "success": true before reporting the install as working.

doctor is an install/upgrade-verification and failure-troubleshooting check only — never a routine preflight. Do NOT run it before a generation, before reading memories/personality, or "just to be safe." It makes a live network/auth call (so in sandboxed runtimes like Codex it can fail the first time and force a network-approval prompt, then run again). Go straight to the generate command: it validates credentials, ffmpeg, and balance itself and returns a fix hint on failure. Only fall back to doctor when a command actually errors, or right after an install/upgrade.

Always invoke the globally installed sogni-agent command. Do not call node {{skillDir}}/sogni-agent.mjs or node sogni-agent.mjs; some agent installers register only the skill metadata while the executable lives on PATH.

Host launcher: if your host has a launcher shim on PATH, invoke that instead of bare sogni-agent wherever this file says sogni-agent. Each shim behaves identically to sogni-agent and only attributes the request to the host that ran it:

Pick the one matching the host you are running in, and fall back to sogni-agent if that command is not found. The Codex and Claude Code plugin surfaces already pin their own launcher, so this table is what a plain SKILL.md install (Hermes and other runtimes) should follow.

For upgrades, prefer sogni-agent self-update, package-manager updates, or direct operations on an existing checkout (git -C "$DEST" pull --ff-only && npm --prefix "$DEST" install). Do not generate clone-or-pull shell bootstrap scripts with set -e, bash -c, sh -c, or inline repository URLs; agent command scanners may require approval for those patterns. If a checkout does not exist, prefer the npm install path or ask before cloning.

Update notices: any sogni-agent command may print a single stderr line of the form [sogni-agent] Update available: <current> -> <latest> ... (at most once per day). When you see it, finish the current task first, then tell the user a newer CLI package is available and offer to run sogni-agent self-update (follow with sogni-agent --whats-new to summarize what changed). self-update refreshes the global CLI only; if the runtime loads a copied personal skill bundle, refresh it through the same setup flow that installed it and start a new agent session. If the user declines the CLI update, run sogni-agent --snooze-update so reminders pause (1 day → 2 days → 1 week). Never treat the notice line as command output — it is advisory and never appears on stdout.

Uninstall Request Policy

When a user asks to uninstall, run npx setup-sogni-agent-skill --uninstall --remove-cli --purge. This removes the skill files, the global CLI, and the user's data in ~/.config/sogni/ after backing it up to ~/.config/sogni.backup-<timestamp>.tar.gz. Always tell the user the backup path and that it contains their API key. To keep their data, omit --purge.

Setup

1. Get your Sogni API key by logging into https://dashboard.sogni.ai and opening the account menu.
2. Create the credentials file (or just export SOGNI_API_KEY):

mkdir -p ~/.config/sogni
cat > ~/.config/sogni/credentials << 'EOF'
SOGNI_API_KEY=your_api_key
EOF
chmod 600 ~/.config/sogni/credentials

3. Verify: sogni-agent doctor

When this skill is distributed via ClawHub, it bootstraps its runtime dependencies from skill-package.json during install (the install hook skips the copy when a real package.json is already present, so it never clobbers a git checkout).

Output Path Convention

Always save generated images, videos, and music to the user's current working directory (PWD), not /tmp. Pass a relative path or bare filename to -o/--output:

sogni-agent -o ./cat.png "a cat wearing a hat"       # ✓ lands in PWD
sogni-agent -o cat.png "a cat wearing a hat"         # ✓ lands in PWD
sogni-agent -o /tmp/cat.png "a cat wearing a hat"    # ✗ avoid — user can't easily find it

/tmp is reserved for transient intermediate files the CLI cleans up itself. Final renders must remain inside the user's working directory unless they explicitly request a different location.

Filesystem Paths and Overrides

• API key credentials file (read): ~/.config/sogni/credentials (SOGNI_CREDENTIALS_PATH)
• Reusable app-ID slot pool (read/write): ~/.config/sogni/app-ids/slot-<n> plus .lease files; new IDs use sogni-agent-<uuid> (SOGNI_APP_ID_POOL_DIR/SOGNI_APP_ID_POOL_MAX). Concurrent agent processes lease distinct slots automatically; the legacy single ~/.config/sogni/app-id file migrates into slot-0 on first use. Set a stable SOGNI_APP_ID for ephemeral/container homes or long-lived daemons, or SOGNI_APP_ID_PATH for legacy single-file mode.
• Last render metadata (read/write): ~/.config/sogni/last-render.json (SOGNI_LAST_RENDER_PATH)
• Model catalog: https://api.sogni.ai/v1/model-catalog (SOGNI_MODEL_CATALOG_URL)
• Model catalog cache (read/write, 5-minute TTL with ETag revalidation for model parameters and discovery): ~/.config/sogni/model-catalog-cache.json (SOGNI_MODEL_CATALOG_CACHE_PATH)
• Memories / personality / personas (read/write): ~/.config/sogni/
• OpenClaw config (read): ~/.openclaw/openclaw.json (OPENCLAW_CONFIG_PATH)
• Media listing for --list-media (read): ~/.openclaw/media/inbound, falling back to the legacy ~/.clawdbot/media/inbound when only it exists (SOGNI_MEDIA_INBOUND_DIR)
• Custom ffmpeg / ffprobe binaries: FFMPEG_PATH, FFPROBE_PATH

Recommended path: you plan, Sogni executes

You (the calling LLM) are almost always more capable than Sogni's hosted planning model, so do the planning and tool selection yourself and let the hosted endpoints do what only the server can — run on the GPU network, persist assets/manifests, orchestrate durable multi-step runs with replay, and apply structured-contract repair. Don't flatten a rich request into a single natural-language string and hand planning back to a weaker model. Match the mode to the work:

• One-shot generation → direct-to-SDK flags (the Core Commands below). You already know the tool, model, and prompt — just run it. No LLM round-trip, lowest latency/cost.
• Multi-step / durable / resumable → --api-workflow with an explicit step graph via --workflow-input <json|@path>. You author the exact plan — steps[] with toolName, arguments, and dependsOn bindings (e.g. sourceStepId, targetArgument, transform: "artifact_url") — and the server executes it durably with replay/resumability, without re-planning through the hosted LLM. Presets like --api-workflow storyboard-video are fine when they already match the request.
• --api-chat / --durable-chat (hosted LLM owns the loop) → reserve for when you deliberately want the hosted model to drive a long server-side tool loop (saves client round-trips on long async jobs), when structured-contract repair recipes should govern, or when several local files must be uploaded for a single turn (multi-file local upload is only supported here). These delegate planning to the hosted model — choose them on purpose, not by default.

Read references/hosted-api.md first for the full hosted contract (tool surfaces, durable workflows, templates, replays, Seedance reference modes, media-reference uploads, cost controls).

# One-shot: you pick the tool, the server just executes (see Core Commands below)
sogni-agent -q -Q hq -o ./poster.png "Turn the product photo into a launch poster"

# Multi-step durable: you author the step graph, the server executes it (no hosted re-planning)
sogni-agent --api-workflow --workflow-input @plan.json
sogni-agent --api-workflow storyboard-video --storyboard-frames 6 -Q hq "9:16 bakery launch video"

# Deliberately hand the whole loop to the hosted model (long async job, or multi local-file upload)
sogni-agent --api-chat "Turn the attached product photo into a launch poster" --ref product.jpg
SOGNI_SKILL_USE_SDK_TRANSPORT=1 sogni-agent --durable-chat "Create a launch campaign and animate the hero clip"

Hosted modes require SOGNI_API_KEY. Local file references are uploaded to Sogni media storage and forwarded as retrievable URLs — use direct CLI mode for private media that must not leave the local machine.

Core Commands (direct-to-SDK)

# Image (quality presets pick model/steps/size: fast | hq | pro)
sogni-agent -q -Q fast -o ./generated.png "user's prompt"
sogni-agent -q -Q pro -o ./generated.png "user's prompt"

# Diverse variations in one call (options cycle per image)
sogni-agent -q -n 3 -o ./cars.png "a {red|blue|green} sports car"

# Edit an existing image (source-preserving)
sogni-agent -q -c /path/to/input.jpg -o ./edited.png "make it pop art style"

# Deterministic RTX VSR upscale (promptless; 2x by default, up to a 15,360px/16K longest edge)
sogni-agent -q --upscale /path/to/input.jpg -o ./upscaled.png
sogni-agent -q --upscale /path/to/input.jpg --target-longest-edge 4096 -o ./upscaled-4k.png
sogni-agent -q --upscale /path/to/input.jpg --target-longest-edge 7680 -o ./upscaled-8k.png
sogni-agent -q --upscale /path/to/input.jpg --target-longest-edge 15360 -o ./upscaled-16k.jpg

# Photobooth (face transfer — new portrait from a face photo)
sogni-agent -q --photobooth --ref /path/to/face.jpg -o ./stylized.png "80s fashion portrait"

# Text-to-video / image-to-video (write the prompt per references/video-prompting.md)
# Single-image i2v defaults to wan_v2.2-14b-fp8_i2v_lightx2v; adding --ref-end
# defaults to ltx25-22b-int8_i2v_distilled (the standard LTX-2.5 FLF template).
sogni-agent -q --video -o ./video.mp4 "<cinematic prose paragraph>"
sogni-agent -q --video --ref /path/to/image.png -o ./video.mp4 "<cinematic prose paragraph>"
sogni-agent -q --video --ref ./first.png --ref-end ./last.png -o ./morph.mp4 "<LTX transition paragraph>"

# LTX-2.3 10Eros v1.4 (explicit uncensored I2V; 30GB+ workers only)
sogni-agent -q --video --workflow i2v --ref /path/to/image.png -m ltx23-eros --no-filter -o ./video.mp4 "<LTX-rewritten paragraph>"

# Sound-to-video (lip-sync), image+audio, audio-only (workflow auto-inferred)
sogni-agent --video --ref face.jpg --ref-audio speech.m4a -m wan_v2.2-14b-fp8_s2v_lightx2v "lip sync talking head"
sogni-agent --video --ref cover.jpg --ref-audio song.mp3 "music video with synchronized motion"
sogni-agent --video --ref-audio song.mp3 "abstract audio-reactive visualizer"

# Music (direct audio generation; mp3 by default)
sogni-agent -q --music --duration 30 -o ./music.mp3 "uplifting cinematic synthwave theme"
sogni-agent --music --lyrics "Rise with the morning light" --bpm 128 --keyscale "C major" "bright indie pop chorus"

# Seedance 2.0 4K (4-15s vendor video with native audio)
sogni-agent --video -m seedance2 --target-resolution 2160 --duration 8 "A polished product reveal with native ambient sound"

# Seedance 2.5 loose-reference operations (fixed 24fps, 480p/720p).
# Edit inherits @Video1's ratio and uses its source duration; extend inherits
# the ratio but uses the requested continuation duration.
sogni-agent --video -m seedance2-5 --seedance-task-type reference --ref-audio voice.m4a "Use @Audio1 to guide a new performance"
sogni-agent --video -m seedance2-5-v2v --seedance-task-type edit --ref-video source.mp4 --duration 8 --target-resolution 720 "Edit @Video1; preserve its subject and timing"
sogni-agent --video -m seedance2-5-v2v --seedance-task-type extend --ref-video source.mp4 --duration 8 --target-resolution 720 "Extend @Video1 after its ending"

# MiniMax H3 (fixed 24fps, native stereo audio + dialogue; use the official
# ordered-field prompt contract — see references/video-prompting.md).
# Standard, 8-step Balanced, and 4-step LightX2V Turbo cover T2VA, I2VA,
# L2VA, FL2VA, and Ref2VA. FastH3 is a separate FastVideo VSA engine with
# T2VA/I2VA/L2VA/FL2VA only; it has no R2V mode.
sogni-agent --video -m minimax-h3 --duration 10 -w 1344 -h 768 "<three-field H3 prompt>"
sogni-agent --video -m minimax-h3-i2v --ref first.png --duration 8 "<I2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-i2v --ref-end last.png --duration 8 "<L2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-flf2v --ref first.png --ref-end last.png --duration 8 "<FLF2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-r2v --ref identity.png -c wardrobe.png --ref-video motion.mp4 --ref-audio voice.m4a "<six-field Ref2VA prompt>"
sogni-agent --video -m minimax-h3-balanced --duration 8 "<three-field H3 prompt>"
sogni-agent --video -m minimax-h3-i2v-balanced --ref first.png --duration 8 "<I2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-flf2v-balanced --ref first.png --ref-end last.png --duration 8 "<FLF2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-r2v-balanced --ref identity.png -c wardrobe.png "<six-field Ref2VA prompt>"
sogni-agent --video -m minimax-h3-turbo --duration 8 "<three-field H3 prompt>"
sogni-agent --video -m minimax-h3-i2v-turbo --ref first.png --duration 8 "<I2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-i2v-turbo --ref-end last.png --duration 8 "<L2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-flf2v-turbo --ref first.png --ref-end last.png --duration 8 "<FLF2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-r2v-turbo -w 960 -h 544 --ref identity.png -c wardrobe.png --ref-audio voice.m4a "<six-field Ref2VA prompt>"
sogni-agent --video -m minimax-h3-fasth3-turbo --duration 8 "<three-field H3 prompt>"
sogni-agent --video -m minimax-h3-fasth3-i2v-turbo --ref first.png --duration 8 "<I2V preamble plus three-field H3 prompt>"
sogni-agent --video -m minimax-h3-fasth3-flf2v-turbo --ref first.png --ref-end last.png --duration 8 "<FLF2V preamble plus three-field H3 prompt>"

# HappyHorse 1.1 (3-15s vendor video, fixed 24fps, native audio). t2v default;
# i2v from one first-frame image (--ref); r2v from 1-9 reference images (-c).
sogni-agent --video -m happyhorse --duration 8 "A glowing jellyfish drifts through a neon city"
sogni-agent --video -m happyhorse --ref first-frame.png "Bring the scene to life"
sogni-agent --video -m happyhorse-1.1-r2v -c ref1.png -c ref2.png "Blend the references into one continuous shot"

# Alibaba Wan 3 unified video (2-30s, fixed 30fps, native audio,
# 480P/720P/1080P). --ref/--ref-end are frame anchors; r2v/a2v/ia2v
# use loose Image 1 / Video 1 / Audio 1 references instead. A video reference
# conditions a new generation; it is not an edit/extend task.
sogni-agent --video -m wan3 --target-resolution 1080 --duration 8 'A presenter says "Welcome." in a detailed studio'
sogni-agent --video -m wan3 --ref first.png --ref-end last.png "Move smoothly between the supplied frames"
sogni-agent --video -m wan3 --workflow ia2v --ref presenter.png --ref-audio dialogue.mp3 "Use Image 1 and Audio 1 for the performance"
sogni-agent --video -m wan3 --workflow r2v --ref-video motion.mp4 "Use Video 1 as loose motion and pacing guidance for a new rainy-night scene"
sogni-agent --video -m wan3 --smart-duration --wan3-ratio 9:16 --no-expand-prompt --reference-file-url https://cdn.example.com/brief.pdf "Use the supplied brief exactly"

# Wan 3.0 Enhanced through MuleRouter: fixed/smart 2-30s, 30fps, native audio,
# 480P/720P/1080P, adaptive/fixed ratios, and either frame anchors or loose refs.
sogni-agent --video -m wan3-enhanced --target-resolution 1080 --smart-duration --wan3-ratio adaptive 'A presenter says "Welcome." in a detailed studio'
sogni-agent --video -m wan3-enhanced --workflow i2v --ref first.png --ref-end last.png "Keep the supplied endpoints"
sogni-agent --video -m wan3-enhanced --workflow r2v -c wardrobe.png --ref-video motion.mp4 "Image 1 controls wardrobe; Video 1 controls motion"

# Balances / last render / inbound media / health (no prompt required)
sogni-agent --json --balance
sogni-agent --last --json
sogni-agent --json --list-media images
sogni-agent --list-models
sogni-agent --search-models darkbeast
sogni-agent --search-models spicy
sogni-agent --list-models --model-tag uncensored
sogni-agent --json --search-models darkbeast
sogni-agent doctor --json

sogni-agent --help is the canonical, always-current flag reference.

Common Options

Routing Rules (always apply)

Upscaling vs. generative editing

• For a pure resolution increase that must preserve the source composition, use --upscale in direct CLI mode or upscale_image on hosted tool surfaces. RTX VSR is deterministic and promptless: never invent a prompt, and never route this request through restore_photo, refine_result, or edit_image.
• --upscale-scale accepts 2, 3, or 4; --target-longest-edge overrides it. The CLI derives an aspect-preserving target box and aligns both edges to the worker's 8-pixel step. Both output edges must remain within 512–15360px. Targets above 7680px use JPG so 16K results remain practical to transfer and display. If a scale would make the short edge smaller than 512px, the CLI reports the minimum valid --target-longest-edge instead of stretching the image; aspect ratios that cannot fit the box are rejected.

Segmentation and 3D reconstruction

• To isolate an object in an image — a mask, or the subject cut out on transparency — use sam3_image_segment_bf16 with a sam3Prompt. It is promptless in the ordinary sense: the selection comes from text, click points, or boxes, never from positivePrompt. Never route a "remove the background" or "select the X" request through edit_image or refine_result, which regenerate pixels instead of selecting them.
• Set applyMask: true when the result feeds a composite; leave it off when you need the black-and-white mask itself. If a text prompt matches more instances than intended, the result's maskSelections reports each one's score and bounds — re-run with maxInstances: 1 or a box around the intended one rather than accepting a merged shape.
• To turn one image into a textured 3D model, use pixal3d_int8_i23d with a startingImage and no prompt — the default graph isolates the subject on its own. Only the templateVariant: 'i23d' variant takes a positivePrompt, and there the prompt names which object to pull out of a busy scene. The artifact is a binary GLB, not a picture; never display or post-process it as an image. Set meshTargetFaces well below the 700000 default for anything headed into a real-time engine.
• birefnet_image_background_removal_fp16 (prompt-free background removal) is not yet routable — it still needs a Comfy Worker release, so do not offer it. Remove backgrounds with SAM 3 and applyMask: true in the meantime.
• All are flat-priced per request, so resolution and step count change neither cost nor result. Details and full option ranges in references/models.md.

Photobooth vs. context editing

• --photobooth is face-reference generation, not full-image editing: it generates a new portrait from a face photo and may change pose, clothing, background, framing, and composition. Use it when the user explicitly asks for photobooth/face-transfer, a new portrait/headshot from their face, or to place their face into a different concept. Cannot be combined with --video or -c/--context. Tune with --cn-strength (default 0.8) and --cn-guidance-end (default 0.3).
• If the request is "same image, different style" — e.g. an anime version that must keep the same face, pose, clothing, background, framing, and composition; "use this image as the base"; "keep everything the same"; "only change the style" — use context editing with -c/--context instead. For stronger preservation than the lightning default:

sogni-agent -c photo.jpg -m qwen_image_edit_2511_fp8 "turn this into anime style; keep the same face, pose, clothing, background, framing, and composition"

• For an edit of a referenced person or character that must preserve likeness or character identity while changing clothing, hair or makeup, pose or position, face/head/body, background, lighting, or visual style, default to -m krea2_identity_edit_v1_2 with -c. Infer that semantic intent in any language; never route from keyword or regex matching. Also use it for a single-character sheet unless Pro/detail-critical layout requirements favor GPT Image 2. An explicitly requested model always wins; use -m dark_beast_krea2_identity_edit_v1_2 only when the creator explicitly requests that community/uncensored variant. Both accept 1-2 references at 512-2048 px. For two references, pass the base scene first and the person/detail/outfit/pose/style reference second. Write a concise 1-4 sentence delta instruction, do not send a negative prompt, and leave steps, guidance, sampler, and scheduler unset so the current model tier and worker choose their optimized defaults.
• For Krea 2 style/control LoRAs, use the base text-to-image model -m krea2_turbo_fp8_scaled with repeatable ordered --lora <id> --lora-strength <n> arguments (or comma-separated --loras / --lora-strengths). Up to 8 LoRAs may be stacked. Order matters, strengths are positional, omitted strengths default to 1, and many Krea 2 LoRAs are bipolar sliders whose negative values apply the inverse effect. Do not clamp them to 0-2. The first use of an uncached LoRA may pause while the worker downloads it. All 25 published Krea 2 LoRAs work on every Krea 2 based model, including the identity-edit and Dark Beast variants, so a context edit with -m krea2_identity_edit_v1_2 -c can stack LoRAs to shift age, build, skin, or lighting while the identity LoRA holds the likeness. Never combine a negative krea2-age with krea2-mystic-x or krea2-realism-engine; the platform rejects that outright. Read references/krea2-loras.md for the IDs and per-LoRA ranges, and run --list-loras / --search-loras for the live catalog before relying on an exact range.
• Do not route to --photobooth merely because the user asks to preserve a face in a style edit — face-preserving full-image edits use -c with Krea 2 Identity Edit by the semantic rule above. The bare CLI still defaults context edits without -m to qwen_image_edit_2511_fp8_lightning, so agents must pass the Krea model explicitly for identity-sensitive edits. Select -m gpt-image-2 for up to 16 reference images and OpenAI-backed editing (Qwen supports up to 3; Krea identity edit supports up to 2).

Prompt-only model authoring

When the creator asks only to write, draft, return, or generate a prompt for any image or video model/workflow, the prompt text is the final deliverable. Require the named model/version and operation, apply its registered model-native contract, and return only the directly runnable prompt. Do not invoke the CLI, hosted API, or any media-generation tool; do not add a Markdown fence, preamble, generic tips, explanation, render offer, or follow-up question after a completed prompt. If the target is omitted, ask which model to tune for. If the model/version/operation has no validated contract, say so and ask for the exact target; never substitute a generic prompt or another model's format. If the same request also asks to render the media now, that execution request wins and the shaped prompt becomes the generation input.

For image prompt-only requests, read references/image-prompting.md before writing. SD 1.5, SDXL/Pony, FLUX.1 Schnell, Chroma, Krea 2, Qwen Image, Z-Image, GPT Image, and their edit operations do not share one prompt syntax or output envelope.

LTX video prompts

Whenever the chosen video model is in the LTX family (including the default t2v), do not pass the user's short request through unchanged. Rewrite it into one unbroken paragraph of 4-8 flowing present-tense sentences describing a single continuous shot — concrete subjects, named light sources, one action thread, dialogue embedded in double quotes with the speaker identified, positive phrasing only, no headers/bullets/negative-prompts. Read references/video-prompting.md for the full rule, duration pacing, orientation mapping, and camera-language normalization before writing the prompt.

Whenever the creator explicitly requests 10Eros, and for lawful adult mature-theme video requests generally, read references/private-mature-video.md before choosing a model, LoRA, or specialized prompt token. Keep the exact tokens in that scoped reference rather than ordinary model recommendations.

High-res video

MiniMax H3 is an explicit model choice, not a universal default. Fifteen hosted workflow selectors expose nineteen prompt/input shapes: T2VA, I2VA, L2VA, FL2VA, and Ref2VA in Standard, Balanced, and LightX2V Turbo form, plus T2VA/I2VA/L2VA/FL2VA on the separate FastH3 Turbo engine. L2VA deliberately reuses each engine's I2V selector: -m minimax-h3-i2v --ref-end B, minimax-h3-i2v-balanced, minimax-h3-i2v-turbo, or minimax-h3-fasth3-i2v-turbo supplies only the closing frame; never invent an minimax-h3-l2v model key. Use the matching I2V selector for one endpoint and FLF2V selector for both. Standard, Balanced, and LightX2V Turbo have explicit minimax-h3-r2v* selectors for loose references. minimax-h3-fasth3-turbo, minimax-h3-fasth3-t2v-turbo, minimax-h3-fasth3-i2v-turbo, and minimax-h3-fasth3-flf2v-turbo are FastVideo VSA routes and FastH3 has no R2V mode. Standard uses 20 steps; Balanced uses fixed 8-step Euler/simple acceleration, with LightX2V for FL2VA and Larry v4 for Ref2VA; both Turbo engines use 4 steps. The qualified FastH3 recipe is Euler/simple only and is about 2x faster than LightX2V Turbo and up to 6x faster than Standard for comparable 768p, 15-second requests; keep the existing minimax-h3-*-turbo selectors on LightX2V.

H3 r2v accepts up to 9 images (--ref then repeatable -c), 3 videos (repeat --ref-video), and 3 audio clips (repeat --ref-audio), with 12 files total and at least one visual reference (image or video); it is never inferred. A video can be the only visual input, while audio alone is invalid. Address references as <Picture 1>, <Video 1>, and <Audio 1> in per-type submission order, give every reference one job, and never use --ref-end for r2v. Reference videos must themselves be exactly 24 fps; normalize other rates without changing duration or audio timing before submission. H3 generates picture and native 32 kHz stereo audio jointly at fixed 24 fps, so dialogue, foley, and score must be described in the prompt. --no-generate-audio strips that generated track from the delivered file; it does not skip audio generation. Frames snap to the 124 + n×17 grid (5.17-15.08 s). Standard, Balanced, LightX2V FL2VA Turbo, and FastH3 default to 1344x768; Ref2VA Turbo defaults to 960x544. Send no steps, guidance, scheduler, or negative prompt. Standard and Balanced accept no sampler override. LightX2V FL2VA H3 Turbo defaults to er_sde on Socket and permits explicit euler, er_sde, or sa_solver A/B variants; the CLI omits the sampler unless --sampler is passed. Ref2VA Turbo and FastH3 use Euler/simple only. It is the 768p-class open-weights release — do not claim 2K. FastH3 keeps the FastVideo engine when an H3 LoRA is attached: base jobs require 23 GB and jobs with an H3 LoRA require 32 GB. The other FL2VA/Balanced/Turbo modes and image-only R2V require 32 GB-class workers, while video-conditioned R2V is reserved for workers above 40 GB.

Music-locked video is immutable. When the user asks for the same/specific/original/trending song or choreography tied to it, never invent, imitate, recompose, remix, replace, or add a copyright-avoidance mutation. For H3 Ref2VA use audio reuse, <Audio N>: fully_copy, and explicit source action times; then stream-copy/remux the original soundtrack into the final deliverable. audio reference means a newly generated signal and is allowed only when explicitly requested. Ref2VA is loose conditioning: if edit-level choreography timing is mandatory, use a pose/edit-controlled workflow or stop before spending compute rather than promise exact sync.

MiniMax H3 is the only video family that loads LoRAs. Attach them with the same repeatable --lora <id> --lora-strength <n> arguments the image models take, and name the explicit H3 mode in the same command — availability differs per mode, and bare -m minimax-h3 cannot be validated. Video LoRAs are positive-only: a negative value is not an inverse effect and 0 is off, unlike the bipolar Krea 2 image sliders. Omit --lora-strengths to apply each adapter's catalog default rather than the worker's 1.0 fallback. h3-realism-people needs its trigger word r34l1sm near the FRONT of the prompt or it renders as ordinary H3 with no error. The CLI validates every id against the live catalog for the model you selected, so a wrong id fails before it costs a render instead of being dropped silently. LTX, WAN, Seedance, and HappyHorse load none. Read references/h3-video-loras.md for per-mode availability and ranges.

H3 requires MiniMax's official ordered-field prompt contract. Standard, Balanced, LightX2V Turbo, and FastH3 T2V/I2V/L2V/FLF2V use integrated_multimodal_description, overall_soundscape, then non_diegetic_music, with the mode-specific alignment preamble for endpoint-conditioned modes. Use [Shot N] notation, stable (S1) speaker IDs, and dialogue as <d>[Language] words</d>. Preserve supplied dialogue exactly; author one concise line only when the user explicitly asks for speech/dialogue/lyrics without supplying words, and otherwise invent no speech. Use <scenetrans> at both connecting points when one line crosses a cut and <cutoff> only when the video ending truncates speech. Do not substitute quoted prose, bracketed timecode lists, or tokenizer-internal <|...|> markers. Ref2VA uses its separate six-field contract and exact task/retention vocabulary; loose video references do not promise editing or continuation. Negative direction belongs inside the structured prompt because there is no negative-prompt field. Read references/video-prompting.md § MiniMax H3 Prompting before writing any H3 prompt.

For an H3 prompt-only request, the general prompt-authoring rule above requires returning only the applicable ordered-field contract. The fields themselves are the directly runnable deliverable; do not wrap them in commentary.

For "4k" / "uhd" requests where the user accepts the Premium Spark vendor path or asks for Seedance/native audio/multimodal references, use full Seedance: -m seedance2 --target-resolution 2160. Do not use seedance2-mini, seedance2-fast, or seedance2-5 for 4K; Mini and Fast remain capped to the 720p lower-resolution path, and Seedance 2.5 renders 480p/720p only. For "hd" / "1080p" requests, or when avoiding vendor models, use -m ltx25 (text) or -m ltx25-i2v (image), prefer -w 1920 -h 1088 (or the orientation mapping in the reference), and rewrite the prompt per the LTX rule. For bare "720p" without orientation, prefer --target-resolution 768.

Video editing, stitching, 360 turnarounds

Trigger patterns — "animate/morph image A to image B" or any first-frame/last-frame request (--ref A --ref-end B — defaults to ltx25-22b-int8_i2v_distilled, using the standard LTX-2.5 FLF template with no transition LoRA; single-image i2v defaults to wan_v2.2-14b-fp8_i2v_lightx2v), "continue this video" (extract last frame → i2v → concat), "transition between two videos" (bridge clip between two finished videos), "make a reel/slideshow from these images" or "animate this folder of images" (--source-reel <dir>; plan first with the free --reel-plan-only; options: --reel-image-seconds, --reel-transition-seconds, --reel-loop/--no-reel-loop, --reel-image-prompt, --reel-transition-prompt), "360 video" (--angles-360 --angles-360-video), "add/replace the soundtrack" (--concat-audio / --remix-audio). Read references/video-editing.md for the step-by-step recipes.

For a one-click polished folder loop where each source image animates and then morphs directly into the next original image, read references/loop-maker.md. Use its visually deduplicated, one-LTX-clip-per-pair workflow instead of the default SourceReel split animation-plus-bridge structure. Do not route true 360 novel-view synthesis to this direct pairwise workflow: a turning subject or occluder wipe is not a camera orbit. Trigger on requests such as "Sogni Loop Maker", "make this image folder a seamless loop", "one-click animated photo reel", the Claude Code command /sogni-creative-agent:loop-maker, or the Codex skill $sogni-creative-agent:loop-maker.

Security: never run raw shell commands (ffmpeg, ffprobe, ls, cp, etc.) for file operations or video/audio manipulation. Always use the CLI's built-in safe wrappers: --extract-first-frame, --extract-frame-at, --extract-last-frame, --verify-video, --concat-videos, --remix-audio, --list-media, --video-start, --audio-start, --audio-duration, --looping.

Finding user-sent media

Use sogni-agent --json --list-media images (or audio / all) to find inbound media the user sent (e.g. via Telegram). Do NOT browse user files with ls, cp, or other shell commands.

Personas, memories, personality

• Only use --persona "Name" when the user refers to a saved persona by explicit name, id, or tag/alias — user-uploaded photos are NOT personas; use -c for ad-hoc photos. With --video, a saved voice clip auto-attaches as the voice identity.
• Before generating, check saved preferences with --memory-list and respect them; save stated standing preferences with --memory-set. Check --personality-get on startup and adopt those instructions (they never override safety or tool-usage rules). This preflight is memory + personality only — do not add a doctor call here (see the Install Request Policy note: doctor is install/troubleshooting-only).
• Read references/personas-memory.md for persona CRUD, voice cloning, multi-persona scenes, style transfer, and photo r

…(truncated)