Quickstart Guide

July 21, 2026 · View on GitHub

Note: For more advanced configurations, see the tutorial and options reference.

Feature Compatibility

For the complete and most accurate feature matrix, refer to the main README.

Model Quickstart Guides

ModelParamsPEFT LoRAFull-RankQuantizationMixed PrecisionGrad CheckpointFlow ShiftTwinFlowSelf-FlowLayerSyncRef InputsControlNetSliders†Guide
PixArt Sigma0.6B–0.9B✗✓int8 optionalbf16✓✗✗✓✓✗✓✓SIGMA.md
NVLabs Sana1.6B–4.8B✗✓int8 optionalbf16✓+✓✓✓✓✗✗✓SANA.md
Kwai Kolors2.7B✓✓not recommendedbf16✓✗✗✗✗✗✗✓KOLORS.md
Stable Diffusion 32B–8B✓✓int8/fp8/nf4 optionalbf16✓+✓ (SLG)✓✓✓✗✓✓SD3.md
Flux.18B–12B✓✓*int8/fp8/nf4 optionalbf16✓+✓✓✓✓✗✓✓FLUX.md
Flux.232B✓✓*int8/fp8/nf4 optionalbf16✓+✓✓✓✓✓ opt✗✓FLUX2.md
Flux Kontext8B–12B✓✓*int8/fp8/nf4 optionalbf16✓+✓✓✓✓✓ req✓✓FLUX_KONTEXT.md
Z-Image Turbo6B✓✓*int8 optionalbf16✓✓✓✓✓✗✗✓ZIMAGE.md
Krea2-✓✓*int8 optionalbf16✓+✓✗✗✗✓ opt✗✓KREA2.md
Boogu-Image 0.1-✓✓*fp8 optionalbf16✓✓✗✗✗✓ edit✗✓BOOGU_IMAGE.md
zlab i13B✓✓int8 optionalbf16✓✓✓✓✓✗✗✓ZLAB_i1.md
Ideogram 49B✓✓*fp8 default, nf4 optionalbf16✓+✓✗✗✗✗✗✓IDEOGRAM4.md
ACE-Step3.5B✓✓*int8 optionalbf16✓✓✓✗✓✗✗✓ACE_STEP.md
Chroma 18.9B✓✓*int8/fp8/nf4 optionalbf16✓+✓✓✓✓✗✗✓CHROMA.md
Auraflow6B✓✓*int8/fp8/nf4 optionalbf16✓+✓ (SLG)✓✓✓✗✓✓AURAFLOW.md
HiDream I117B (8.5B MoE)✓✓*int8/fp8/nf4 optionalbf16✓✓✓✓✓✗✓✓HIDREAM.md
OmniGen3.8B✓✓int8/fp8 optionalbf16✓✓✗✓✗✗✗✓OMNIGEN.md
Stable Diffusion XL2.6B✓✓not recommendedbf16✓✗✗✗✓✗✓✓SDXL.md
Lumina22B✓✓int8 optionalbf16✓✓✓✓✗✗✗✓LUMINA2.md
Cosmos22B✓✓not recommendedbf16✓✓✓✓✓✗✗✓COSMOS2IMAGE.md
Cosmos316B-65B✓✓*no_change firstbf16✓✓✗✗✗audio opt✗✓COSMOS3.md
LTX Video~2.5B✓✓int8/fp8 optionalbf16✓✓✓✓✓✓ I2V✗✓LTXVIDEO.md
LTX Video 219B✓✓*int8/fp8 optionalbf16✓✓✓✓✓✓ opt✗✓LTXVIDEO2.md
Hunyuan Video 1.58.3B✓✓*int8 optionalbf16✓✓✓✓✓✓ I2V✗✓HUNYUANVIDEO.md
Wan 2.x1.3B–14B✓✓*int8 optionalbf16✓✓✓✓✓✗✗✓WAN.md
Wan 2.2 S2V14B✓✓*int8 optionalbf16✓✓✓✓✓✗✗✓WAN_S2V.md
Qwen Image20B✓✓*required (int8/nf4)bf16✓✓✓✓✓✗✗✓QWEN_IMAGE.md
Qwen Image Edit20B✓✓*required (int8/nf4)bf16✓✓✓✓✓✓ req✗✓QWEN_EDIT.md
Stable Cascade (C)1B, 3.6B prior✓✓*not supportedfp32 (required)✓✗✗✗✗✗✗✓STABLE_CASCADE_C.md
Kandinsky 5.0 Image6B (lite)✓✓*int8 optionalbf16✓✓✓✓✗✓ I2I✗✓KANDINSKY5_IMAGE.md
Kandinsky 5.0 Video2B (lite), 19B (pro)✓✓*int8 optionalbf16✓✓✓✓✓✓ I2V✗✓KANDINSKY5_VIDEO.md
LongCat-Video13.6B✓✓*int8/fp8 optionalbf16✓+✓✓✓✓✓ opt✗✓LONGCAT_VIDEO.md
LongCat-Video Edit13.6B✓✓*int8/fp8 optionalbf16✓+✓✓✓✓✓ req✗✓LONGCAT_VIDEO_EDIT.md
LongCat-Image6B✓✓*int8/fp8 optionalbf16✓✓✓✓✓✗✗✓LONGCAT_IMAGE.md
LongCat-Image Edit6B✓✓*int8/fp8 optionalbf16✓✓✓✓✓✓ req✗✓LONGCAT_EDIT.md

✓ = supported, ✓ = requires DeepSpeed/FSDP2 for full-rank, ✗ = not supported, ✓+ indicates checkpointing is recommended due to VRAM pressure. Ref Inputs marks existing reference/edit/I2V conditioning paths; opt means optional, req means the edit/I2V flavour requires it. TwinFlow ✓ means native support when twinflow_enabled=true (diffusion models need diff2flow_enabled+twinflow_allow_diff2flow). Self-Flow ✓ means native support for crepa_enabled=true with crepa_feature_source=self_flow, use_ema=true, and crepa_teacher_block_index set. LayerSync ✓ means the backbone exposes transformer hidden states for self-alignment; ✗ marks UNet-style backbones without that buffer. †Sliders apply to LoRA and LyCORIS (including full-rank LyCORIS "full"). All models support LyCORIS.*

ℹ️ Wan quickstart includes 2.1 + 2.2 stage presets and the time-embedding toggle. Flux Kontext covers editing workflows built atop Flux.1.

⚠️ These quickstarts are living documents. Expect occasional updates as new models land or training recipes improve.

Fast paths: Z-Image Turbo & Flux Schnell

  • Z-Image Turbo: Fully supported LoRA with TREAD; runs fast on NVIDIA and macOS even without quant (int8 works too). Often the bottleneck is just trainer setup.
  • Flux Schnell: The quickstart config handles the fast noise schedule and assistant LoRA stack automatically; no extra flags needed to train Schnell LoRAs.

Advanced Experimental Features

  • Diff2Flow: Allows training standard epsilon/v-prediction models (SD1.5, SDXL, DeepFloyd, etc.) using a Flow Matching loss objective. This bridges the gap between older architectures and modern flow-based training.
  • Scheduled Sampling: Reduces exposure bias by letting the model generate its own intermediate noisy latents during training ("rollout"). This helps the model learn to recover from its own generation errors.

Common Issues

Dataset has fewer samples than expected

If your dataset ends up with fewer usable samples than you expected, files may have been filtered during processing. Common reasons include:

  • Files too small: Images below minimum_image_size are filtered out
  • Aspect ratio out of range: Images outside minimum_aspect_ratio/maximum_aspect_ratio bounds are excluded
  • Duration limits: Audio/video files exceeding duration limits are skipped

Viewing filtering statistics:

  • In the WebUI, browse to your dataset directory and select it to see filtering statistics
  • Check the logs during dataset processing for statistics like: Sample processing statistics: {'total_processed': 100, 'skipped': {'too_small': 15, ...}}

For detailed troubleshooting, see Troubleshooting filtered datasets in the dataloader documentation.