EXPERIMENTAL — Mage-Flow Character LoRAMage-Flow support in ComfyUI is only days old as of this writing, and getting a working generation out of it required tracking down and fixing several non-obvious bugs. Marking this experimental accordingly, and documenting exactly what it took below so you don't have to repeat the same debugging.HOW IT WAS TRAINEDTrained with ai-toolkit against the Mage-Flow-Base diffusers checkpoint (via the SceneWorks/Mage-Flow-Base mirror), 3000 steps, using ai-toolkit's built-in mageflow extension. Standard character-LoRA setup, no unusual training parameters. All 22 dataset images were generated by Grok Imagine, based on the description of an original character from the book "And All His Heart's Desires."HOW IT WAS TESTEDTested in SwarmUI + ComfyUI, using ComfyUI's native Mage-Flow support (merged upstream in PR #15026, plus the bf16 fix in PR #15081). Getting a real image out of this pipeline — rather than a solid-noise "watered silk" pattern — required working through three separate bugs stacked on top of each other:1. Model misclassification: if your ComfyUI/SwarmUI install had already scanned the checkpoint file before Mage-Flow support existed, it may have cached the file as generic Qwen-Image and won't re-check unless the file's modified time changes. If your model loads but text-encoder shapes mismatch (2560 vs 3584), touch the checkpoint file and restart your backend.2. FP8 quantization breaks Mage-Flow: loading the diffusion model with fp8_e4m3fn weight precision silently collapses output to noise on current ComfyUI/PyTorch builds. Use bf16 ("16-bit"/default precision), not fp8.3. Scheduler matters: the "simple" scheduler produces garbage/noise-like output for this model. Use "linear_quadratic" instead.CONFIRMED WORKING SETTINGSBase model (mage_flow_edit_base_bf16):- Precision: bf16 (not fp8) · Sampler: euler · Scheduler: linear_quadratic · Steps: 30 · CFG: 5 · Sigma Shift: 6 · VAE: mage_flow_vae_bf16 · Text Encoder: Qwen3-VL-4B · LoRA weight: 1.0Turbo model (mage_flow_edit_turbo_bf16): - Same precision/scheduler/VAE/encoder · CFG: ~1 (guidance-distilled — real CFG over-guides and fries the output) · Steps: 6 (cleaner hands/skin than the "standard" 4, minimal speed cost) · Sigma Shift: 6ONE MORE THING: you may see a "not compatible with current model" warning on this LoRA. Cosmetic only — ai-toolkit writes the wrong architecture tag (qwen-image/lora instead of mage-flow/lora) even though the trained weights correctly target Mage-Flow's layers. Safe to ignore.

Files with SHA256: fd65a693d643cb0b08d94cd7479ee98265fd3ee6f40588aeee1f0882a93b11fd