huggingface / huggingface/diffusers
Multi-controlnet batching for `StableDiffusionXLControlNetInpaintPipeline`
- Vorherrschende Sprache
- Python
- Sterne
- 34.5k
- Forks
- 7.3k
- Ø Merge
- 3 T. 3 Std.
- Gemergte PRs (30 T.)
- 91
Beschreibung
**Is your feature request related to a problem? Please describe.**
Currently, batching is not supported when we are conditioning the SDXL pipeline on multiple controlnets: https://github.com/huggingface/diffusers/blob/28f9d84549c0b1d24ef00d69a4c723f3a11cffb6/src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py#L771
**Describe the solution you'd like.**
We can pass a list of lists of images for batch inference.
**Describe alternatives you've considered.**
One-by-one processing, which results in very low GPU utilization.
There was a similar PR raised for the `StableDiffusionControlNetPipeline` https://github.com/huggingface/diffusers/pull/6334
Beitragsleitfaden
Rechercherichtung
Start in src/diffusers/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py at the referenced batching logic. Compare the requested list-of-lists input with the similar StableDiffusionControlNetPipeline change in PR #6334. Done means multi-ControlNet SDXL inpaint inference accepts batched inputs and avoids one-by-one processing.
Vom Indexierungsmodell aus dem Issue-Text verfasst.
Bewertung
- Tech-Stack
- python, pytorch
- Bereich
- machine-learning
- Issue-Typ
- Feature
- Schwierigkeit
- 4/5
- Geschätzter Aufwand
- 3-5 Tage
- Aktivitätsstatus
- Ruhig
- Klarheit
- Größtenteils klar
- Anfängerfreundlichkeit
- 52/100