Extensions/krea2-pose-control-LoRA
ComfyUI Extension

krea2-pose-control-LoRA

Skeleton-conditioned Control-LoRA for Krea-2. Combines flow-matching MSE with Keypoint R-CNN pose-consistency loss to drive precise pose transfer across visual domains. Trained on Krea-2 Raw, deployed on Krea-2 Turbo.

By s-adhit·Created 3 months ago·Updated a day ago· 3
s-adhit/krea2-pose-control-LoRA
Nodes—
On cloudLocal install
Stars3
Updateda day ago
Readme

Krea-2 Pose Control-LoRA

<p align="center"> <img src="https://raw.githubusercontent.com/s-adhit/krea2-pose-controlnet/main/docs/showcase/final/hero-v2/final/final_generation_montage.png" alt="Krea-2 Pose Control final generations" width="100%"> </p> <p align="center"> <img src="https://raw.githubusercontent.com/s-adhit/krea2-pose-controlnet/main/docs/showcase/final/hero-v2/final/final_condition_montage.png" alt="Krea-2 Pose Control pose conditions" width="100%"> </p>

Frozen pose-conditioned image generation for Krea-2 Turbo. A rendered PoseBridge skeleton controls spatial human geometry; the prompt controls compatible appearance and rendering.

The released control mechanism is not a parallel ControlNet branch. The clean pose latent is spatially aligned and channel-concatenated with the noisy image latent before an expanded ControlInputLayer. Krea-2 Raw was used for training; the image-input backbone remains frozen while the control half of that layer and rank-64 LoRA adapters are trained. Frozen inference uses Krea-2 Turbo.

Frozen release

  • Artifact: krea2-pose-control-mix025.safetensors
  • Release identity: krea2-pose-control-lora-v1, candidate mix-025
  • SHA-256: 6d97e9c2e102e07928fc8864346401a0d2e6082d610ca6b037c4704102e3f8d1
  • Defaults: Turbo, 8 steps, CFG 0, mu=1.15, control scale 1.0, native/aspect-preserving geometry, no Style-LoRA

The release decision, artifact lineage, locked runtime, and limitations are in docs/evaluation/release/final_release_v1.json.

Architecture

rendered skeleton -> clean pose latent
noisy image latent + clean pose latent -> channel concat -> ControlInputLayer
                                                     -> frozen Krea-2 + rank-64 LoRA

The prompt must not compete with the condition on person count, framing, or geometry. See prompting.md for the control/prompt contract.

Evidence and assets

Training and evaluation are complete. The primary result record is docs/blog-evidence/EVALUATION_RESULTS.md; methodology and caveats are beside it in docs/blog-evidence/. Frozen benchmark specifications and payloads live in docs/evaluation.

Qualitative and blog-ready assets are indexed in docs/blog-assets/BLOG_ASSET_INDEX.md, including the hero showcase, geometry comparison, hard-pose examples, and the appendix regeneration manifest.

Inference and ComfyUI

Use inference.py for the canonical CLI contract. It validates the frozen release before loading and writes an output provenance sidecar. The CLI defaults to native/aspect-preserving geometry; explicit dimensions must be positive multiples of 16.

comfyui/README.md documents the completed standalone ComfyUI integration, including the DWPose-to-PoseBridge conversion path and frozen model paths.

Repository status

Model development, training, evaluation, frozen release, and ComfyUI integration are complete. The repository is in its blog/documentation and evidence-archive phase; it does not have a future adapter roadmap.

References