Blankage
patreon
April Request #1 - Wan2.1 TG/RC Test
🕑 Added 2025-04-27 18:33:08 +0000 UTCFinal keyframe was a woman and I image2imaged backwards to the man using the BigAsp 2.0 checkpoint. I was very pleased with how well this worked overall. Still, I wouldn't consider these optimal keyframes and I was very impressed with how well Wan2.1 handled them.
I used the model on Fal.ai via the API because you can't disable the safety checker otherwise. It might be possible to improve on the results using a comfyui workflow, but I need to figure out the best hyperparameters. I'll be testing that shortly!
I joined two clips, but the prompt for the first clip was "一段逼真的手机视频，一位男士低头看着手中的文件，他的胸部越来越丰满，也越来越女性化。他留着一头黑色的短发，逐渐变长。他坐在床上，身后是一堵白墙。镜头慢慢拉远，展现出他日益女性化的身影。特写，低头"
I used the standard negative in Chinese: "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
I read somewhere that the "first-to-last-frame" model was trained on Chinese, and it did appear to work better than with English (based on some quick comparisons I did). Of course, I'm still learning how to prompt this model so I may be prompting the wrong way entirely.
Comments
W0RR13Z
Also apologies for being confusing
Blankage
Glad you like it! I don't think it's an exaggeration to say this is the best result I've ever gotten with an open source model lol. But yeah, I hope it will be better with higher-quality keyframes. What do you mean by "looking for voice work"?
W0RR13Z
I like this a lot, seems a bit rough but I’m sure with some time you’ll grasp the program and prompts. You looking for Voice work? If not no worries but I do like the concept!
