Same face, new pose: give it one reference image of your character and get her back in a different pose, with the face, hairstyle and skin tone kept. Free ComfyUI workflow, one JSON file, no LoRA and no training.
Use it with an adult character you generated yourself, or a photo you have the rights to. Do not use it on photos of real people who have not agreed to it.
One reference image on the left, the same character in a new seated pose on the right.
WHAT YOU GET
• Face and look kept from your reference image (one character per image)
• 12 ready-made poses: standing, seated, walking, half-body, front / side / back turnaround, close-up, and four relaxed poses
• A text box for your own pose or framing
• 2 outfit levels: 1 keep the clothes, 2 modest everyday outfit
• SeedVR2 upscale 2x built in (switch it off with one tick)
• Random seed each run, written into the file name
Pose 3 (walking, outfit level 1) and pose 2 (seated, outfit level 2).
Pose 4 (half-body) and pose 6 (side view), outfit level 2.
Pose 1 (standing), pose 11 (hip tilt) and pose 12 (seated, legs crossed), outfit level 2. Only poses that were checked are shown; poses 5, 7, 8, 9 and 10 were not tested.
HOW TO USE
1) Drag the workflow JSON into ComfyUI.
2) Group 1: pick your own reference image.
3) Group 2: type outfit level 1 or 2. Group 3: type pose 1 to 12.
4) Group 4: a short scene and light sentence (default: soft daylight, plain background).
5) Press Run (older ComfyUI builds call it Queue).
MODELS (not included, download them yourself)
• qwen_image_2.1_int8_convrot.safetensors (7.26 GB) → ComfyUI/models/diffusion_models/
• qwen3vl_8b_int8_convrot.safetensors (9.35 GB) → ComfyUI/models/text_encoders/
• qwen_image_2.1_vae_bf16.safetensors (0.68 GB) → ComfyUI/models/vae/
• seedvr2_ema_7b_fp16.safetensors (16.5 GB) and ema_vae_fp16.safetensors (0.50 GB) → ComfyUI/models/SEEDVR2/ (the SeedVR2 node can also download them itself)
Links:
• https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors
• https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/seedvr2_ema_7b_fp16.safetensors
• https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/ema_vae_fp16.safetensors
CUSTOM NODES (ComfyUI Manager: "Install Missing Custom Nodes")
• ComfyUI-SeedVR2_VideoUpscaler (tested 2.5.24): https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler
• ComfyUI-Impact-Pack (tested 8.28.3): https://github.com/ltdrdata/ComfyUI-Impact-Pack
• ComfyUI_essentials (tested 1.1.0): https://github.com/cubiq/ComfyUI_essentials
• rgthree-comfy (tested 1.0.2608210019): https://github.com/rgthree/rgthree-comfy
Tested on ComfyUI 0.43.0. The two Qwen 2.1 nodes need a recent ComfyUI; the minimum version was not checked.
SEEDVR2 UPSCALE
Same face region before and after the 2x SeedVR2 step (same image, same seed). The face stays the same; skin and hair get finer detail.
SPEED AND VRAM
Measured on one machine only: a 16 GB card, 14,946 to 14,980 MiB peak (about 15 GB) with SeedVR2 on, 45 to 82 seconds per image over 5 test images (the first image is the slowest). Landscape reference → 2752x1536 output, portrait 3:4 reference → 1536x2028. With less VRAM, switch SeedVR2 off. Other cards were not measured.
LIMITATIONS
• Tested with one character and one reference image; other faces and styles may not hold.
• The poses shown in the example images were checked by eye; other poses may need a few seeds.
• Pose 12 with a landscape reference is cut at mid-thigh; use a portrait reference.
• Edit sizes above 1024 were not tested.
Built with Qwen. Qwen-Image 2.1 is licensed under the Qwen Research License Agreement (research and evaluation use; commercial use needs a separate license from the licensor): https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE
Upscaling by SeedVR2 (ByteDance; ComfyUI integration by numz and contributors).