Call for training data (and how to contribute)
☕ Optional: Support on Ko-fi — all money goes to compute.
Call for training data
This is a solo hobby project and right now I hand-author all the training data myself in my spare time, which is the main thing limiting how fast the model improves. More good data means better results, directly. If you would like to help push future versions forward, contributed datasets are hugely welcome, exactly like stablellama's head/face/eye/person swap dataset that made those capabilities possible in v1.2.
What helps most (roughly in order):
- Character reference sheets — multiple consistent views/expressions of one character. Biggest current gap.
- Two-person / group scenes with distinct, clearly different identities (helps identity separation).
- Different camera angles of the same subject or scene (front, side, three-quarter).
- Full-body restaging where the face is small in frame.
- Same edit phrased many different ways (helps the model stop being picky about wording).
Format: edit triples of source image + instruction + target image, where everything outside the requested edit stays identical between source and target. stablellama's dataset is a good structural reference. Reference sheets can be sets of consistent-character images.
Do not let format stop you. If all you have is reference sheets or consistent images of a character with no edits, that is still a big help — we can generate the target edits from them ourselves. Share whatever you have, in whatever form it exists. Getting it into training shape is our job, not yours.
Two rules, please:
- License it permissively (MIT or similar) so it can actually be used.
- Only contribute data you have the rights to. Prefer AI-generated or clearly consented material. No non-consensual or real-person data without consent, in line with the responsible-use note on this model.
Publish your dataset on Hugging Face, drop a link in this thread, and I will credit every contributor whose data ends up in a release. Thank you.
Contributors
Everyone whose data or work has helped so far, thank you:
- stablellama — head/face/eye/person swap dataset (v1.2)
- CeciliaXCIX — community support
- akashzeno — node engineering
- SubtleShader — training-code testing
- LPX55 — offered a large synthetic multi-scene likeness dataset
- Kimboooo — offering character reference sheets
- Mark (sogni.ai) — support, GPU-fund donation, and spreading the word
Your name goes here if your contribution lands in a release.
hello,有示例数据集格式以供参考吗?我不确定应该提供哪些数据集是符合要求的\高质量的
We have some incredibly hq datasets that would be incredibly helpful for this (wide array of poses, actions, scenes, etc. all synthetic but 4K and likeness created in 3, 4, and 6 scenes.) Unfortunately we have scrapped the project after spending hundreds of combined hours, as our focus is urgently elsewhere at the moment. We would rather keep it private for the time being, so publishing the dataset is a no from us, but definitely open to sharing as long as you keep it private. Feel free to shoot me an email if you want a sample. They're not in the format that you requested, so you must also be prepared to implement some sort of image transformation pipeline and dataset organization.
@LPX55 :you can upload data to Huggingface in a private dataset and then grant individually rights to access. So you can easily use this as a transfer without granting public access.
This is amazing, thank you. Genuinely grateful you would share it even after moving on to other things, and please do not spend any more effort than is easy for you. Whatever you can share, in whatever format it already exists, we will take it from there. Getting it into training shape and figuring out how to use it is our problem, not yours.
A private Hugging Face dataset with access granted to me works perfectly (thanks @stablellama for the idea), or email if that is easier. I have added you to the contributors shout-out list. Thank you again, this is exactly the kind of thing that keeps the project moving.
Thank you for wanting to help. The only thing we really need is the core triple: an input image, an instruction, and a target image, where everything outside the edit stays the same between the two. stablellama's dataset is a good structural reference: https://huggingface.co/datasets/stablellama/change_eye_face_head_person . Do not worry about matching an exact format or structure though. Share whatever you have in whatever form it exists, and we handle the formatting, organization, and cleanup on our end.