Rebels Krea2 Outpaint

ComfyUI-Rebels-Krea2-Outpaint

<img width="49%" alt="ident_test__00011_" src="https://github.com/user-attachments/assets/57373d2c-323b-4ce4-b001-8fba982c4d67" /> <img width="49%" alt="krea2_outpaint_00006_" src="https://github.com/user-attachments/assets/4e198ecf-7ec5-4ad5-bff5-4a41848f7fa2" />

Registered outpainting for the yijunwang2/krea2-outpaint LoRA on Krea 2 Turbo, in ComfyUI.

This is a ComfyUI port of that LoRA's diffusers pipeline. The LoRA is not a plain edit LoRA: the source image is registered into the target canvas grid at an explicit bounding box, so the model knows where the known pixels belong. This pack does exactly that registration and nothing else exotic — the known region is carried only by the reference, and the exact source pixels are restored at the end with a feathered seam.

What it is under the hood

The reference-attention machinery (t=0 reference modulation, isolated-ref K/V precompute, frame axis i+1) is the same mechanism ostris ships in ComfyUI-Krea2-Ostris-Edit. The one and only difference for outpaint is the reference RoPE coordinates: instead of the reference getting its own y/x grid from (0,0), its tokens are mapped into the target latent grid at the bbox with center-sampled fractional spacing. That mirrors _pack_reference_latents in the outpaint pipeline exactly. The reference still lives on RoPE frame i+1 — it is not coplanar with the target.

Requirements

• A ComfyUI build with native Krea 2 support (or your GGUF Krea 2 fork — see caveat below).
• Krea 2 Turbo diffusion model + the Krea 2 VAE + the Krea 2 Qwen3-VL CLIP.
• The LoRA weights krea2_outpaint_rank32.safetensors in models/loras.
• https://huggingface.co/yijunwang2/krea2-outpaint/blob/main/krea2_outpaint_rank32.safetensors
• No extra Python dependencies.

Install:

cd ComfyUI/custom_nodes
git clone https://github.com/RealRebelAI/ComfyUI-Rebels-Krea2-Outpaint
# restart ComfyUI

Nodes (category: RealRebelAI/krea2_outpaint)

• Rebels Krea2 Outpaint Canvas — source + direction + target W/H → condition, placed_source, placement (JSON), canvas_width, canvas_height.
• Rebels Krea2 Outpaint Encode — clip + prompt + vae + condition → CONDITIONING (reference latent attached). vlm_reference is OFF by default, matching the LoRA's encode_reference_in_prompt=False contract.
• Rebels Krea2 Outpaint Model Patch — model + placement + kv_cache → patched MODEL. Place it after Load LoRA. Keep kv_cache ON (the LoRA uses isolated reference attention).
• Rebels Krea2 Outpaint Composite — generated + placed_source + placement → final image with exact known pixels restored.

Wiring

Load Diffusion Model (krea2 Turbo)
    -> Load LoRA (krea2_outpaint_rank32)
    -> Rebels Krea2 Outpaint Model Patch (placement, kv_cache=True)
    -> KSampler

CLIPLoader (krea2) -> Rebels Krea2 Outpaint Encode (prompt, vae, condition) -> KSampler.positive
CLIPLoader (krea2) -> CLIP Text Encode ("")                                 -> KSampler.negative

Rebels Krea2 Outpaint Canvas:
    condition       -> Encode.condition
    placement       -> Model Patch.placement  AND  Composite.placement
    placed_source   -> Composite.placed_source
    canvas_width/height -> EmptyLatentImage

EmptyLatentImage (canvas W/H) -> KSampler.latent
KSampler -> VAE Decode -> Rebels Krea2 Outpaint Composite -> Save Image

Sampler settings (Turbo)

• steps 8, sampler euler, CFG 1.0 (Krea convention guidance 0.0 → CFG 1.0 in ComfyUI, which also skips the uncond pass), denoise 1.0.
• The empty latent is the full canvas — pure noise, no latent masking. The reference does the anchoring.
• Prompt should describe the complete output image, not just the new area.

Direction options

extend_right, extend_left, extend_down, extend_up, extend_width_both, extend_height_both. The node builds a bbox that preserves the source aspect ratio and spans the full complementary dimension (single pass). It raises with a clear message if there's no room to extend in that direction at the chosen canvas size.

Notes / limitations

• Single pass only in this version — the source box spans the full width or full height. That covers extend-in-one-direction and centered-both-sides. Fully interior boxes (canvas on all four sides) need the two-pass plan; the planner is included (plan_passes) and the placement JSON reports pass_count, but chaining the two sampler passes is manual for now.
• GGUF fork caveat: the model patch reaches into the native Krea 2 SingleStreamDiT internals (dit.first, dit.blocks, dit.pe_embedder, dit.patch, block attn/mod, etc.). A GGUF loader that swaps the linear ops but keeps that module tree should work unchanged; if your fork renames or restructures those attributes, the patch is where to adjust.
• Rectangular outpainting only — this is not a mask-based inpainter.

Credits & licenses

• Reference-attention machinery adapted from ostris/ComfyUI-Krea2-Ostris-Edit (MIT, © 2026 Ostris, LLC).
• Registered placement geometry and pipeline math from the yijunwang2/krea2-outpaint release (pipeline.py, outpaint.py, Apache-2.0; portions © 2026 Ostris, LLC and Krea AI / HuggingFace).
• This pack's node code: MIT, © 2026 RealRebelAI. See LICENSE and NOTICE.

The LoRA weights themselves are under the Krea 2 Community License; commercial misuse should be reported to Krea ([email protected]).