[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-92579":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":15,"stars30d":16,"stars90d":15,"forks30d":15,"starsTrendScore":15,"compositeScore":17,"rankGlobal":9,"rankLanguage":9,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":9,"pushedAt":9,"updatedAt":23,"readmeContent":24,"aiSummary":25,"trendingCount":15,"starSnapshotCount":15,"syncStatus":14,"lastSyncTime":26,"discoverSource":27},92579,"comfyui-krea2edit","lbouaraba\u002Fcomfyui-krea2edit","lbouaraba","Instruction-based, identity-preserving image editing for Krea 2 in ComfyUI — nodes + workflows for the Krea 2 Identity Edit LoRA",null,"Python",128,6,3,2,0,75,50.04,"Apache License 2.0",false,"main",true,[],"2026-07-22 04:02:06","# ComfyUI-Krea2Edit\n\nInstruction-based image editing for **Krea 2** in ComfyUI — the node pack that powers\nthe **Krea 2 Identity Edit** LoRA. Turns Krea 2 (Raw or Turbo) into an image editor with dual\nconditioning: the source image is injected both as VAE latent tokens (appearance) and\ninto the Qwen3-VL text encoder (semantic grounding), matching how the LoRA was trained.\n\n## Model versions\n\nSee [CHANGELOG.md](CHANGELOG.md) — **v1.1 is recommended** (better likeness,\nlocality, and remove\u002Freplace; two honest caveats listed there).\n\n## Installation\n\n```bash\ncd ComfyUI\u002Fcustom_nodes\ngit clone https:\u002F\u002Fgithub.com\u002Flbouaraba\u002Fcomfyui-krea2edit\n# restart ComfyUI\n```\n\nRequirements: a ComfyUI version with native Krea 2 support, the Krea 2 model\n(Raw or Turbo), the Qwen3-VL 4B text encoder used by Krea 2, and the Krea 2 Identity Edit\nLoRA (`krea2_identity_edit_v1.safetensors`). No extra Python dependencies.\n\n## Nodes\n\n### `Krea2EditModelPatch`\nWraps the diffusion model so the VAE-encoded source image is prepended as clean\nin-context tokens (RoPE frame 1). Inputs:\n- `model` — Krea 2 (LoRA already applied)\n- `source_latent` — VAEEncode of the image being edited\n- `source_latent_b` *(optional)* — second reference (RoPE frame 2) for two-input\n  edits (e.g. person + scene)\n\n### `Krea2EditGroundedEncode`\nImage-grounded instruction encoding — the text encoder *sees* the image while\nreading your instruction, exactly as during training. Inputs:\n- `clip` — the Krea 2 CLIP (Qwen3-VL, loaded with `type: krea2`)\n- `prompt` — the edit instruction (\"recolor the car to matte black\")\n- `image` — the same source image\n- `image_b` *(optional)* — second reference for two-input edits\n- `grounding_px` — grounding resolution (default 768; trained range 512–1536).\n  This is a quality dial: lower = stronger edit adherence, higher = stronger\n  identity\u002Flikeness. Try 1024+ for people, 512 for stubborn scene changes.\n\n**Both nodes are required.** With a stock `CLIPTextEncode` the model never sees the\nimage semantically and quality drops sharply, especially for scene-referential\ninstructions (\"the man on the left\").\n\n## Minimal wiring\n\n```\nLoadImage ─┬─ VAEEncode ── Krea2EditModelPatch.source_latent\n           └─ Krea2EditGroundedEncode.image     (+ your prompt)\nUNETLoader ── LoraLoaderModelOnly (krea2_identity_edit_v1 @1.0) ── Krea2EditModelPatch.model\nKrea2EditModelPatch ── KSampler.model\nKrea2EditGroundedEncode ── KSampler.positive\nKrea2EditGroundedEncode (empty prompt, same image) ── KSampler.negative\nEmptySD3LatentImage ── KSampler.latent_image\n```\n\nExample workflows in `workflows\u002F`: `krea2_edit_single_ref.json`, `krea2_edit_two_ref.json`.\n\n## Usage notes (read these — they matter)\n\n1. **Match the aspect ratio.** The target latent (`EmptySD3LatentImage`) must have\n   the same aspect ratio as the source image. Training pairs are same-size; a\n   mismatched AR is out of distribution and visibly degrades identity\u002Fpreservation.\n2. **Turbo, 8 steps, CFG 1** is the fast path (~1 min at 2MP) and works for most\n   edits: recolor, add\u002Finsert, attribute changes, restyles, scene translation.\n3. **Removals and other \"delete salient content\" edits need real guidance:**\n   use the **Raw** model at **CFG 3, ~20 steps**. Distilled Turbo at CFG 1 will\n   usually re-render the subject instead of removing it.\n4. At CFG > 1, ground the negative too: a second `Krea2EditGroundedEncode` with an\n   empty prompt and the same image (this is the trained unconditional).\n5. Two-input edits: scene image → `source_latent`\u002F`image`, subject image →\n   `source_latent_b`\u002F`image_b`. Leave the b-inputs unconnected for single-image use.\n6. **Generate at ≤2MP.** Above the trained range, source content can bleed into\n   the output or subjects duplicate.\n7. **Two people with distinct faces:** chain single-ref inserts (place person A,\n   then run a second edit adding person B from their reference) — currently more\n   face-faithful than one two-ref pass.\n\n## License \u002F credits\n\nNodes: Apache-2.0. The **Krea 2 Identity Edit** weights ship separately under the\nKrea 2 Community License Agreement (see the model card, `LICENSE.pdf`, and `NOTICE`\nin the weights repo).\nBuilt on Krea 2 by Krea AI; text encoder Qwen3-VL (Alibaba).\n","这是一个为ComfyUI设计的Krea 2图像编辑插件，支持基于自然语言指令的身份保持图像编辑。核心功能包括双路径条件注入：源图像既作为VAE潜在表示参与扩散过程，又通过Qwen3-VL文本编码器实现语义对齐；提供Krea2EditModelPatch和Krea2EditGroundedEncode两个必需节点，确保指令理解与视觉内容强关联。适用于人像编辑、局部重绘、风格\u002F属性修改等需高保真身份保留的场景，尤其适合Krea 2 Raw\u002FTurbo模型与Identity Edit LoRA协同使用。","2026-07-09 02:30:23","CREATED_QUERY"]