Startup LoRA fusion for FLUX.1-dev

The model runner accepts one LoRA repository or local file for FLUX.1-dev:

xdit --model black-forest-labs/FLUX.1-dev \
  --lora_path alvdansen/frosting_lane_flux \
  --lora_weight_name flux_dev_frostinglane_araminta_k.safetensors \
  --lora_scale 0.75 \
  --prompt "A small robot, frstingln illustration"

The runner uses Diffusers/PEFT to load the adapter and safely merge it into the base weights. This fusion runs before device placement, compilation, or caching. The runner then removes the adapter modules, so generation has no separate LoRA computation. Fusion changes weights in memory and preserves the downloaded base checkpoint.

This support is specific to the FLUX.1-dev runner and to MiniMax-H3 Turbo. For FLUX.1-dev, quantized loading, FSDP, meta loading, and PipeFusion are rejected before loading weights. Restart the runner to change or remove the adapter; startup fusion cannot switch adapters during execution. The legacy pipeline example CLI does not expose these options.

--lora_weight_name is optional when Diffusers can select the weight file. --lora_scale defaults to 1.0 and must be finite. A scale of zero preserves the base weights. LoRA weights must be compatible with the selected base model.

MiniMax-H3 Turbo

MiniMax-H3-Turbo runs the base MiniMax-H3 checkpoint with a distilled step-reduction LoRA from lightx2v/Minimax-h3-Turbo fused into the transformer at startup. Without --lora_path it downloads the FL2VA Turbo 8-step v1.0 768p adapter and samples 8 steps with video shift 6, the runner's default adapter settings:

xdit --model MiniMax-H3-Turbo --task i2va \
  --ulysses_degree 8 --gemm_quantization fp8 \
  --input_images first_frame.png --prompt "..."

--lora_path selects another adapter: a safetensors file, a directory, or a repository, with --lora_weight_name naming the file in the latter two. --lora_scale multiplies the weight update. The adapter scaling factor alpha comes from the file's metadata. If it is missing, the runner uses 8, as in v0.1. For the adapters in the model specs, --num_inference_steps and --flow_shift default to the published values and other step counts are refused:

Adapter (--lora_weight_name) Steps Video shift
minimax_h3_fl2v_turbo_8step_v1.0_768p_bf16.safetensors (default) 8 6
minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors 4 6
minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors 8 (or 4) 12
minimax_h3_fl2v_turbo_4step_v0.1.safetensors 4 12

The 544p adapters were trained at about 0.5 megapixels. Set a matching resolution, such as --height 544 --width 960, instead of the 768×1344 default. For any adapter outside the table, set --num_inference_steps and --flow_shift. This includes later releases in the repository. The audio shift stays at the checkpoint's 3.0 for all of them.

Fusion runs before FP8 conversion, so MiniMax-H3-Turbo accepts supported quantized GEMMs. Do not combine it with FSDP, memory-efficient meta loading, or PipeFusion. Startup fails if an adapter tensor has no matching module, lacks its A or B half, or has the wrong shape. The runner rejects Ref2VA and ComfyUI-format adapters. The base MiniMax-H3 and FastH3 runners do not accept --lora_path.