Just a real latest image edit models compare
retained text
Just a real latest image edit models compare, before you drop more shit on Flux 3 Klein.
Same seed, no picking.
Inspiration to do this? I make character sheets and refmods and wanted to check Fibo as latest model. Sure, with custom loras results will change, but I compare pure models here.
Same prompt, just adjusted for specific model image tags, like <ref\_image\_0> for F3K. Model files used are all int8.
No custom loras, models set to their usual samplers and steps; qwen 2.1 full + official turbo lora. No "upsampler" for F3K.
https://preview.redd.it/4umf8y8bwrwh1.jpeg?width=5024&format=pjpg&auto=webp&s=bf617bf7340b37ea2aa16ec3a448bd759dc35a83
Redraw character from <image2> and create a photorealistic character using reference images. Make rear view.
Use <image1> exclusively as the reference for character's pose, camera angle and viewing direction and composition.
Use <image2> exclusively as the reference for the character's identity and appearance.
Reproduce the exact same character shown in <image2>. Preserve facial identity, facial structure, bone proportions, eyes, nose, lips, hairstyle, hairline, hair color, hair texture, skin tone, skin undertone, and apparent age.
Character's outfit is determined exclusively by:
<image3>: upper-body garment (top).
Reproduce item from <image3> accurately, preserving its original design, colors, materials, texture and details.
The character has balanced natural proportions. Use realistic human anatomy.
Photorealistic studio photography, neutral studio lighting, and a pure white background.
And... F3K - only model that not only was able to recognize pose but also show jacket from correct side.
document fetched 2026-10-22 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xbgtdn/just_a_real_latest_image_edit_models_compare