Segue a descrição em inglês, pronta para colar no Civitai.<hr /><h1 id="krea-2-identity-edit-a+b-pose-control-batch-candidates-selective-refine">Krea 2 Identity Edit — A+B · Pose Control · Batch Candidates · Selective Refine</h1>Restage a person from a single reference photo, with a consistent face and identity, in any scene, outfit or lighting. You can optionally add a second character and lock the pose with a depth reference.The workflow runs in two passes. First it generates several candidates quickly. Then it refines only the one you pick, so you don't spend upscale time on images you won't keep.<h2 id="features">Features</h2><ul><li>Identity from 1 photo. Uses the Krea 2 Identity Edit LoRA with grounded Qwen3-VL encoding and the in-context source patch.</li><li>Optional 2nd character (Identity B). One toggle turns it on or off.</li><li>Optional pose control. A depth map from any photo locks the composition and pose. It affects only the composition stage, so identity is not affected.</li><li>Batch candidates. Generate N variations, then refine only the chosen one. Stage 1 is reused from cache.</li><li>Final refine. Upscale ×1.5, a low-denoise polish pass and a light sharpen.</li><li>One-click toggles. REFINE / 2nd CHARACTER / POSE. Bypassed groups need no input image.</li></ul><h2 id="how-it-works">How it works</h2><ol><li>Composition (steps 1–2): text and pose only, so the layout stays free from the reference.</li><li>Identity (steps 3–8): the identity patch and grounded encode bring in the face and appearance.</li><li>Refine (selected image only): upscale, a polish pass at 0.17 denoise, and sharpen.</li></ol><h2 id="how-to-use">How to use</h2><ol><li>Load your reference photo in IDENTITY A.</li><li>Write the prompt: scene, pose, outfit and lighting.</li><li>Turn REFINE off, set <code>batch_size</code>, and Run. Candidates are saved as <code>krea2_candidate_*</code>.</li><li>In PICK → batch_index, choose your favorite (0 = first).</li><li>Turn REFINE on and Run again. Only the refine stage runs.</li><li>For a new set of candidates, change the seed and keep it on <code>fixed</code>.</li></ol>2nd character: turn the toggle on, load the second photo, and describe both people in the prompt, for example "the woman from the first reference and the man from the second reference…".Pose: turn the toggle on and load a vertical (9:16) pose photo. Keep <code>amount</code> equal to <code>batch_size</code>.<h2 id="defaults-and-vram">Defaults &amp; VRAM</h2><ul><li>Base resolution is 1.0 MP with batch 2. The refine upscale is ×1.5, for about 2.25 MP.</li><li>Low VRAM: use 0.5 MP, upscale ×2, and batch 1.</li><li>Two characters plus pose plus batch is the heaviest combination. If you get OOM or tensor errors, set batch 1, set the seed to <code>increment</code>, and Run several times.</li></ul><h2 id="requirements">Requirements</h2>Models<ul><li><code>krea2_turbo_int8_convrot.safetensors</code> (or the fp8 version), placed in diffusion_models</li><li><code>qwen3vl_4b_fp8_scaled.safetensors</code>, placed in text_encoders</li><li>VAE: <code>qwen_image_vae.safetensors</code> / <code>wan_2.1_vae.safetensors</code></li><li>LoRA: <code>krea2_identity_edit_v1_2.safetensors</code> (required)</li><li>LoRA: <code>realism_engine_krea2_v2.safetensors</code> (optional)</li><li>Upscaler: <code>2xNomosUni_span_multijpg_ldl.pth</code></li><li>Pose: <code>depth-control-lora.safetensors</code>, from Patil/Krea-2-depth-controlnet</li></ul>Custom nodes<ul><li>comfyui-krea2edit</li><li>comfyui-krea2-conditioning</li><li>comfyui-krea2-controlnet</li><li>comfyui_controlnet_aux</li><li>RES4LYF</li><li>rgthree-comfy</li><li>ComfyUI-Easy-Use</li><li>KJNodes</li><li>WAS Node Suite</li><li>Comfyroll</li><li>comfyui-various</li><li>ComfyUI-Inspyrenet-Rembg</li><li>Krea2 T-Enhancer</li></ul><h2 id="tips">Tips</h2><ul><li>Pose not taking: raise Control LoRA strength to 1.0. If it copies the silhouette too much, lower it to 0.6.</li><li>Pose still drifting: raise <code>steps_to_run</code> on the first sampler from 2 to 3.</li><li>Over-saturated look: lower the Rebalance strength (default 1.75).</li><li>Two-image mode: the LoRA was trained as scene + person. Two people works, but likeness is a bit weaker than with one reference.</li></ul>

Krea2 Identity Restage | 1-2 Characters · Depth Pose · Batch Refine

Krea2_IdentityEdit_AB_Pose_Batch_Refine_V2.json

v1.0

Segue a descrição em inglês, pronta para colar no Civitai.<hr /><h1 id="krea-2-identity-edit-a+b-pose-control-batch-candidates-selective-refine">Krea 2 Identity Edit — A+B · Pose Control · Batch Candidates · Selective Refine</h1>Restage a person from a single reference photo, with a consistent face and identity, in any scene, outfit or lighting. You can optionally add a second character and lock the pose with a depth reference.The workflow runs in two passes. First it generates several candidates quickly. Then it refines only the one you pick, so you don't spend upscale time on images you won't keep.<h2 id="features">✨ Features</h2><ul><li>Identity from 1 photo. Uses the Krea 2 Identity Edit LoRA with grounded Qwen3-VL encoding and the in-context source patch.</li><li>Optional 2nd character (Identity B). One toggle turns it on or off.</li><li>Optional pose control. A depth map from any photo locks the composition and pose. It affects only the composition stage, so identity is not affected.</li><li>Batch candidates. Generate N variations, then refine only the chosen one. Stage 1 is reused from cache.</li><li>Final refine. Upscale ×1.5, a low-denoise polish pass and a light sharpen.</li><li>One-click toggles. REFINE / 2nd CHARACTER / POSE. Bypassed groups need no input image.</li></ul><h2 id="how-it-works">🧩 How it works</h2><ol><li>Composition (steps 1–2): text and pose only, so the layout stays free from the reference.</li><li>Identity (steps 3–8): the identity patch and grounded encode bring in the face and appearance.</li><li>Refine (selected image only): upscale, a polish pass at 0.17 denoise, and sharpen.</li></ol><h2 id="how-to-use">🚀 How to use</h2><ol><li>Load your reference photo in IDENTITY A.</li><li>Write the prompt: scene, pose, outfit and lighting.</li><li>Turn REFINE off, set <code>batch_size</code>, and Run. Candidates are saved as <code>krea2_candidate_*</code>.</li><li>In PICK → batch_index, choose your favorite (0 = first).</li><li>Turn REFINE on and Run again. Only the refine stage runs.</li><li>For a new set of candidates, change the seed and keep it on <code>fixed</code>.</li></ol>2nd character: turn the toggle on, load the second photo, and describe both people in the prompt, for example "the woman from the first reference and the man from the second reference…".Pose: turn the toggle on and load a vertical (9:16) pose photo. Keep <code>amount</code> equal to <code>batch_size</code>.<h2 id="defaults-and-vram">⚙️ Defaults &amp; VRAM</h2><ul><li>Base resolution is 1.0 MP with batch 2. The refine upscale is ×1.5, for about 2.25 MP.</li><li>Low VRAM: use 0.5 MP, upscale ×2, and batch 1.</li><li>Two characters plus pose plus batch is the heaviest combination. If you get OOM or tensor errors, set batch 1, set the seed to <code>increment</code>, and Run several times.</li></ul><h2 id="requirements">📦 Requirements</h2>Models<ul><li><code>krea2_turbo_int8_convrot.safetensors</code> (or the fp8 version), placed in diffusion_models</li><li><code>qwen3vl_4b_fp8_scaled.safetensors</code>, placed in text_encoders</li><li>VAE: <code>qwen_image_vae.safetensors</code> / <code>wan_2.1_vae.safetensors</code></li><li>LoRA: <code>krea2_identity_edit_v1_2.safetensors</code> (required)</li><li>LoRA: <code>realism_engine_krea2_v2.safetensors</code> (optional)</li><li>Upscaler: <code>2xNomosUni_span_multijpg_ldl.pth</code></li><li>Pose: <code>depth-control-lora.safetensors</code>, from Patil/Krea-2-depth-controlnet</li></ul>Custom nodes<ul><li>comfyui-krea2edit</li><li>comfyui-krea2-conditioning</li><li>comfyui-krea2-controlnet</li><li>comfyui_controlnet_aux</li><li>RES4LYF</li><li>rgthree-comfy</li><li>ComfyUI-Easy-Use</li><li>KJNodes</li><li>WAS Node Suite</li><li>Comfyroll</li><li>comfyui-various</li><li>ComfyUI-Inspyrenet-Rembg</li><li>Krea2 T-Enhancer</li></ul><h2 id="tips">💡 Tips</h2><ul><li>Pose not taking: raise Control LoRA strength to 1.0. If it copies the silhouette too much, lower it to 0.6.</li><li>Pose still drifting: raise <code>steps_to_run</code> on the first sampler from 2 to 3.</li><li>Over-saturated look: lower the Rebalance strength (default 1.75).</li><li>Two-image mode: the LoRA was trained as scene + person. Two people works, but likeness is a bit weaker than with one reference.</li></ul>

Files with SHA256: 27a544630fb3c318a3a98f25033f645455a203681aa58bf4f132904f5f5af008