Towards dgenerate 5.1

[iwr-redmond]

Happy new year @Teriks! I hope your vacation continues to be enjoyable and that you have avoided the cold weather.

I reckon it would be good to take a snapshot of FRs for the next dgenerate version, which I have called 5.1 for convenience. 5.0 was a massive undertaking, so hopefully 5.1 can be more incremental. However, there are still some good features waiting for you to consider in due course.

Issues Summary

Feature Request	Group	Status	Complexity	Notes
#72 SD2 Repository Change	Chores	New	Easy
#76 Deprecate Stable Cascade	Chores	New	Easy	Depends on diffusers>=0.36.0
#17 PuLID Face Transfer	Generation Enhancements	Postponed	Moderate	Needs testing to determine minimum VRAM required
#47 Regional Guidance	Generation Enhancements	Holdover	Hard
#48 MuLan Language Adapter	Generation Enhancements	Postponed	Moderate	Needs testing with 1B LLM for reduced memory usage
#68 Euler Scheduler Variants	Generation Enhancements	New	Easy	Ported from k-diffusion by @R3gm
#70 DyPE Flux Hi-Resolution	Generation Enhancements	New	Moderate	May have memory spike issues
#28 Generation Preview	Library Features	Holdover	Easy / Moderate	Easy if torch-jpeg works as hoped
#62 Detect Missing VAE	Library Features	Holdover	Moderate
#69 Scripting Image Variables	Library Features	New	Easy
#67 SDNQ Version Bump	Memory Efficiency	New	Easy	Now available on PyPI
#77 Control LoRAs	Memory Efficiency	New	Easy	Depends on diffusers>0.36.0
#78 Group Offloading	Memory Efficiency	New	Moderate	May be incompatible with some pipelines
#79 Aule Attention	Memory Efficiency	New	Moderate	Good for cross-platform speed-ups
#83 Bitsandbytes for XPU & MPS	Memory Efficiency	New	Moderate
#13 o1 Pipelines	New Architectures	Holdover	Hard	Integrate k-diffusion into SD 1x pipelines
#19 Hunyuan-DiT	New Architectures	Postponed	Easy
#27 Fooocus SDXL Modules	New Architectures	Postponed	Hard	Porting Fooocus features to Diffusers
#60 Framepack F1 Video Generation	New Architectures	Holdover	Moderate
#73 xllamacpp Prompt Upscaler	New Architectures	New	Moderate
#56 SDXL T5+CLIP Pipeline	New Pipelines	Postponed	Easy
#58 Flux Kontext Inpainting	New Pipelines	Holdover	Moderate	May not be feasible
#61 Pix2Pix Inpainting Pipelines	New Pipelines	Holdover	Moderate	May not be feasible
#80 DDUF Single Files	Offline Inference	New	Easy
#52 Vendored Config Files	Offline Inference	Holdover	Easy	Continuing process
#44 CPU-Only Inference	Speed Enhancements	Holdover	Easy	Potential for extra speed gains with #67 and #74
#75 Cache-DiT	Speed Enhancements	New	Easy	HiDiffusion-like enabling mechanism
#82 Kontext RegionE	Speed Enhancements	New	Easy	May not work well with auto-upscaling

My suggestion for 5.1 is to focus on the chores, library features, offline inference, and speed enhancement groups, as well as bumping SDNQ to its new PyPI version (#67) because it can be managed normally along with other dependencies once this is done. You may also wish to consider adding in xllamacpp (#73) to significantly boost the number of LLMs that can be used for scripting.

[iwr-redmond]

@Teriks:

I've been given access to cto.new, which means I could throw its AI programmer at some of these FRs if you are okay with reviewing PRs coauthored by AI. LMK what you think when you are next online.