huggingface / huggingface/diffusers
Various Chroma pipeline issues
- Dominant language
- Python
- Stars
- 34.5k
- Forks
- 7.3k
- Avg merge
- 3d 3h
- Merged PRs (30d)
- 91
Description
### Describe the bug
Some minor issues that I have found while using the Chroma pipeline for the first time:
- if `sentencepiece` is not installed, you get a quite cryptic error message. Other pipelines such as Flux give you a nice error message that `sentencepiece` has to be installed
- `The config attributes {'guidance_embeds': False, 'pooled_projection_dim': 768} were passed to ChromaTransformer2DModel, but are not expected and will be ignored. Please verify your config.json configuration file.`
Chroma does not use CLIP, only T5, so there is no pooled projection. and does not use a guidance embedding as Flux - but I guess this is an issue with the config file, not the diffusers pipeline
- this code https://github.com/huggingface/diffusers/blob/03c3f69aa57a6cc2c995d41ea484195d719a240a/src/diffusers/pipelines/chroma/pipeline_chroma.py#L806 seems unnecessary, because dynamic timestep shifting is disabled in Chroma's scheduler config: https://huggingface.co/lodestones/Chroma1-HD/blob/main/scheduler/scheduler_config.json
confirmed by the actual timesteps that are used:
`(Pdb) timesteps
tensor([1000., 975., 950., 925., 900., 875., 850., 825., 800., 775.,
750., 725., 700., 675., 650., 625., 600., 575., 550., 525.,
500., 475., 450., 425., 400., 375., 350., 325., 300., 275.,
250., 225., 200., 175., 150., 125., 100., 75., 50., 25.],
device='cuda:0')`
- Chroma's official repo is `lodestones/Chroma1-HD` now. This could be updated in your pipeline example code: https://huggingface.co/docs/diffusers/main/api/pipelines/chroma
### Reproduction
run the sample code here https://huggingface.co/docs/diffusers/main/api/pipelines/chroma
### Logs
```shell
```
### System Info
diffusers HEAD, python 3.11.11
### Who can help?
_No response_
Contributor guide
Assessment
This issue has not been assessed yet.