Download Latest Version v4.9.2 - minor bugfixes and QoL source code.zip (11.4 MB)
Email in envelope

Get an email when there's a new version of SimpleTuner

Home / v4.8.0
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-08-21 4.2 kB
0
v4.8.0 - InfiniteTalk, training metrics UI, new distillers and representation learning methods source code.tar.gz 2026-08-21 17.6 MB
0
v4.8.0 - InfiniteTalk, training metrics UI, new distillers and representation learning methods source code.zip 2026-08-21 20.0 MB
1 1 weekly downloads
Totals: 3 Items   37.6 MB 1

Features

  • InfiniteTalk now supported with VRAM-vs-Performance tuned example configurations provided
  • Cubic spline distribution for flow-matching timesteps
  • Built-in tensorboard / W&B replacement with Prometheus exporter and webUI components (Metrics page)
  • iREPA spatial representation alignment with a training-only module added to capture this information
  • Self-Transcendence distillation, which doesn't reduce step count or bake in CFG, but instead teaches earlier layers in the model things that tend to be learnt by the later layers instead, possibly reducing redundancy and making inner layers "less redundant"
  • Internal guidance training, similar to Self Transcendence except its module is used at inference time; it can bake CFG into the finetune, but requires inference code tweaks to be useful
  • MixFlow, which is comparable to SimpleTuner's own ReflexFlow implementation, aims to reduce exposure bias in flow matching models by teaching them to correct errors they've made earlier in their trajectory without changing inference code requirements or slowing training down (wall clock per step)
  • DiffusinonBlocks totally changes the training objective for a model to reorganise its internals into specialised denoising experts; inference time absolutely needs the same changes AND routing config to be enabled

Bugfixes

  • audio_only is automatically enabled for audio datasets on video models
  • Multi-caption inputs no longer clobber other text embeds' cache on keyed cache models (eg. image/video models that embed the image/video into the text embeds; Qwen Edit, Krea2 edit, and other similar edit models)
  • MultiGPU context-parallel DDP batches are generated deterministically to ensure the batch always shared correctly
  • Audio VAEs now function with vae_batch_size>1 without crashing
  • Live dataset config is now authoritative, emitting a better error when changing immutable properties (not allowed to be changed mid-training / during resume)
  • H3 context-parallel and audio validation have minor fixes

What's Changed

New Contributors

Full Changelog: https://github.com/bghira/SimpleTuner/compare/v4.7.1...v4.8.0

Source: README.md, updated 2026-08-21