HOW TO REALIFY IMAGES WITHOUT CHANGING SHAPE/EXPRESSION?
retained text
Hi everyone! I'm currently working with a dataset containing multi-view facial images (some examples attached), and my goal is to make these images more photorealistic while maintaining **cross-view consistency**. However, there's one major constraint: **the facial geometry, expressions, and landmark positions must remain unchanged.** I already have dense facial landmarks associated with these images, so any geometric distortion introduced during the transformation could affect my downstream experiments. I've experimented with ChatGPT's image generation capabilities, and the results have been promising in terms of realism. However, I'm looking for a more efficient, scalable, and reproducible workflow that can process the entire dataset while preserving the original facial structure. I've come across approaches involving **Anime2Real** and **Qwen-Image-Edit**, but I'd like to get some community feedback before committing to a particular pipeline. A few things I'm particularly interested in: * **Geometry preservation:** How can I improve photorealism without altering facial shape, expressions, or dense landmark positions? * **Multi-view consistency:** How can I ensure that identity, textures, and facial details remain consistent across different camera angles? * **Efficiency:** Are there existing models, workflows, or conditioning techniques that would work well for large-scale dataset generation? I have access to substantial university computing resources, so computational requirements aren't a major limitation. However, with my thesis deadline approaching, I'd prefer an existing solution or a workflow that doesn't require extensive training from scratch. Any suggestions would be greatly appreciated! Thanks in advance!
document fetched 2026-11-15 00:29 · canonical https://reddit.com/r/StableDiffusion/comments/1xvz8qd/how_to_realify_images_without_changing