Krea

Krea

Krea ships two completely separate things under one brand, and picking the wrong one wastes the job.

The fork, decide it first.

• Hosted API. Nodes Krea2ImageNode and Krea2StyleReferenceNode, category partner/image/Krea, billed per image, capped at 1K output, and the only path that reads a Krea moodboard from the website.
• Open weights. Krea 2 Raw and Turbo run locally on core nodes, free, any resolution, no moodboard.

Who owns what, so you open one file, not three. THIS file owns the fork, the hosted graph, the FLUX.1 Krea Dev graph and the ecosystem map. reference.md next to it owns full node I/O, the price matrix, weight files with measured sizes, and the custom-node packs. The kit's MODELS.md already owns the local Krea 2 graph in depth (Raw vs Turbo settings, the qwen3vl_4b encoder and its krea2 CLIP type, the Ostris edit method, the style-reference LoRA on core nodes, the nine official style LoRAs, the abliterated-encoder swap, licence). Do not restate it here, route to it.

Which Krea for which job

The hosted graph, buildable

Two nodes, one optional chain. Confirmed against the shipped templates api_krea2_t2i and api_krea2_style_reference on 2026-08-06.

                    LoadImage -> Krea2StyleReferenceNode ---(KREA_STYLE_REF)--+
                    LoadImage -> Krea2StyleReferenceNode <--(chain in)--------+
                                          |
                                          +--(KREA_STYLE_REF)-> Krea2ImageNode -> SaveImage

Style references chain node to node: each Krea2StyleReferenceNode takes an optional incoming style_reference and appends one more, so the last node in the chain is the one you wire to Krea2ImageNode. Ten is a hard ceiling and the node raises rather than truncating.

Krea2ImageNode widget order, which is what you fill when you build the graph in code:

[0] prompt   [1] model   [2] aspect_ratio   [3] resolution
[4] creativity   [5] moodboard_id   [6] moodboard_strength   [7] seed   [8] control_after_generate

The official template ships "Krea 2 Medium", "1:1", "1K", "medium", "", 0.35.

Three model tiers, and the source states what each is for: Medium for expressive illustrations, Large for expressive photorealism, Medium Turbo for speed. Prices differ 4x between Turbo and Large, so the tier is a budget decision as much as a look decision. Full matrix in reference.md.

Things that will bite you here:

• creativity is a combo of raw / low / medium / high, not a number. The embedded node docs shipped with ComfyUI say FLOAT, which is wrong; they carry their own "AI-generated" disclaimer. The node source wins.
• resolution accepts only 1K. There is no 2K option to find.
• moodboard_id must be a real UUID copied from the Krea website or the node raises before spending anything. Only one moodboard per request.
• Style-reference strength runs -2.0 to 2.0, and negative values invert the style influence rather than disabling it. Set 0 to neutralise, not a negative.
• Billing follows the most expensive feature you switched on: a moodboard costs more than a style reference, which costs more than plain text, and setting both charges the moodboard rate.

FLUX.1 Krea Dev, buildable

Black Forest Labs' Krea-tuned FLUX, aimed at a photographic look with less of the plastic "AI sheen". It is an ordinary FLUX dev graph, which is the whole point: no new node types.

UNETLoader(flux1-krea-dev_fp8_scaled.safetensors, default)
DualCLIPLoader(clip_l.safetensors, t5xxl_fp16.safetensors, type=flux) -> CLIPTextEncode -> KSampler.positive
                                                            -> ConditioningZeroOut -> KSampler.negative
EmptySD3LatentImage(1024, 1024, 1) -> KSampler.latent_image
KSampler(steps 20, cfg 1.0, euler, simple, denoise 1.0) -> VAEDecode <- VAELoader(ae.safetensors) -> SaveImage

Confirmed by reading the official template flux1_krea_dev.json and expanding its subgraph, 2026-08-06. Note what is absent: the template carries no FluxGuidance node at all, and the negative branch is a ConditioningZeroOut off the same encode. Verified by string search on the template, not by eye. If you add FluxGuidance you are departing from the shipped recipe, so change one thing and compare.

Weights: Comfy-Org/FLUX.1-Krea-dev_ComfyUI, one file, split_files/diffusion_models/flux1-krea-dev_fp8_scaled.safetensors, 11.90 GB (11.09 GiB) measured, ungated. The original black-forest-labs/FLUX.1-Krea-dev is gated, so reach for the repack unless you need the reference weights. Licence flag: flux-1-dev-non-commercial-license.

Krea Realtime 14B, and why ComfyUI is not the road most travelled

Real, open, Apache-2.0. A ComfyUI pack exists but is unproven; the paths with users behind them are elsewhere.

• Distilled from Wan-AI/Wan2.1-T2V-14B with Self-Forcing, which converts a diffusion video model into an autoregressive one. Their own figure is 11 fps at 4 steps on a single B200, with roughly one second to the first frame. Text to video and video to video, including webcam and canvas input, with the prompt changeable mid-generation.
• The parts that are the actual contribution: KV Cache Recomputation and KV Cache Attention Bias to hold off error accumulation over a long autoregressive roll.
• There IS a ComfyUI pack, and it is unproven. eliteprox/ComfyUI-Krea ships KreaRealtimeVideoLoader and KreaRealtimeVideoGenerate and loads the 14B checkpoint. 0 stars, last push 2025-11-04. Treat it as a lead, not a route, and read its source before wiring anything. This entry previously said no ComfyUI node existed. That was wrong: the search behind it looked for the checkpoint FILENAME in code across GitHub and this pack does not name it the same way. A negative claim needs several differently-shaped searches before it is worth stating.
• What to run instead: their own FastAPI server (github.com/krea-ai/realtime-video, web app on port 8000), DiffSynth-Studio, or LightX2V.
• Fit on a 24 GB card: the official single file is 28.58 GB (26.61 GiB) measured, so it does not fit. The community fp8 repack 6chan/krea-realtime-video-fp8 is 14.29 GB (13.31 GiB) in both e4m3fn and e5m2, which does. That repack makes no ComfyUI claim and the card is frontmatter only.

The Krea 2 custom-node ecosystem worth knowing

Core plus what MODELS.md documents covers text to image, the Ostris edit path and the style-reference LoRA. These three fill jobs none of that covers. Adoption numbers read on 2026-08-06; none of the three ships a licence file.

• lbouaraba/comfyui-krea2edit (over 400 stars, last push 2026-07-29). Instruction editing that holds a face, built around the Krea 2 Identity Edit LoRA. Node Krea2EditModelPatch injects the source image twice, as VAE latent tokens for appearance and through the Qwen3-VL encoder for semantics, which is how the LoRA was trained. Takes a second reference for person-plus-scene edits. Its trainer is public too.
• facok/comfyui-krea2-controlnet (162 stars, 2026-07-04). Depth, canny, pose and lineart control for Krea 2 through a Control LoRA: Krea2 Control LoRA Loader -> Krea2 Control Image Encode (feed it a comfyui_controlnet_aux preprocessor output) -> Krea2 Control Apply -> sampler. It fails loudly when the loader runs without an attached control latent instead of silently sampling half-patched, which is the behaviour you want.
• Per-layer conditioning control, and it is a genuine two-horse choice. The upstream is nova452/Rebalance-Pack (formerly ComfyUI-ConditioningKrea2Rebalance) (477 stars, pushed 2026-07-29), which introduced reweighting Krea 2's twelve Qwen3-VL conditioning taps. The fork huwhitememes/comfyui-krea2-conditioning (126 stars, 2026-06-26) argues the upstream multiplies the whole tensor by 4 on top of the per-layer gains, inflating conditioning magnitude by roughly 8.7x, and RMS-renormalises instead so the ratios shift while total magnitude holds. The upstream is the more adopted of the two by a wide margin, so the fork is a minority position that makes a technical case, not the consensus. Reach for either when a Krea 2 job shows skin artifacts or likeness drift; nobody here has run an A/B.

Not adopted and not recommended yet, listed so nobody re-finds them and assumes they are missing knowledge: comfyui-krea2-nunchaku, comfyui-krea2-sageattention-guard, comfyui-krea2-fal-lora-converter, comfyui-krea2-controlnetPlus, ComfyUI_Krea_Nodes. All at 0 to 2 stars.

When it goes wrong