I am having issues with the Ref2VA face swap; what am I doing wrong?
retained text
I’m trying to replace the woman in a 10-second video using four identity photos. Ref2VA completes the render, but the result **keeps the original woman** from the driving video almost unchanged. Setup Model: minimax\_h3\_ref2va\_pruned\_int8\_convrot Text encoder: Heretic INT8 Workflow: MiniMaxH3ReferenceToVideo Inputs: four separate target-identity images plus one complete driving video Resolution: 736×736 Length: 243 generated frames, trimmed to 240 frames/10 seconds FPS: 24 Steps: 8 Sampler: MiniMaxH3TurboSampler Reference image size: max Video shift: 12 Audio shift: 3 Optimization: memory-sparse No SAM/mask, face-swap adapter, or additional character LoRA Prompt Use <Video 1> as the complete full-frame motion, timing, pose, camera, clothing, anatomy and room composition. Recreate the woman throughout the entire video with the same adult identity shown consistently in <Picture 1>, <Picture 2>, <Picture 3> and <Picture 4>. Preserve her distinctive face shape, green eyes, nose, lips, warm skin tone and long dark hair from every angle while adapting them naturally to each head pose and expression in <Video 1>. Keep the original body movement, hands, clothing, camera and background coherent. Render every frame as one seamless full image without a pasted face, crop boundary, mask seam, duplicated body region or localized patch. The four pictures contain only the target identity. The original performer appears only in <Video 1>. Is the full reference video overpowering the identity images because Ref2VA treats it as appearance conditioning rather than motion-only conditioning? Should I use different prompt tags/settings, a corrected first frame, fewer video-reference frames, pose extraction, or a separate tracked face-swap stage? Attachments 1. Original driving video 2. Target identity sheet or the four individual crops 3. Generated Ref2VA output **What am I doing wrong?**
document fetched 2026-09-20 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1wkui41/i_am_having_issues_with_the_ref2va_face_swap_what