π New FL2V & REF2V Turbo LoRAs Released β 4/8 Steps, BF16 + ComfyUI
We've released updated MiniMax-H3-Turbo LoRAs! π
New checkpoints
| Model | BF16 | ComfyUI |
|---|---|---|
| FL2V 4-step v1.3 | BF16 | ComfyUI |
| FL2V 8-step v1.1 | BF16 | ComfyUI |
| REF2V 4-step v1.0 | BF16 | ComfyUI |
| REF2V 8-step v1.1 | BF16 | ComfyUI |
Recommended FL2V inference setting
For the new FL2V 4-step and 8-step LoRAs, we recommend using a video flow shift of 6 during inference.
In LightX2V, set video_flow_shift to 6.0. This recommendation applies to FL2V, not REF2V.
Please use the inference step count matching your checkpoint.
We'd love to see your results and hear your feedback. Thanks for your support! β€οΈ
NiceοΌοΌοΌ
Ref 4 step just makes blurry jittery visual nonsense with nonsense prompt following with refmods.
Ref 8 step makes incoherent speech and bad prompt following with refmods.
That's the extent of the feedback i can give from my experience.
In contrast this prior version worked pretty well with same setup: minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors
what video flow shift? any sample comfy workflow?
thanks
Possible regression in FL2V 4-step v1.3 with VEDA, H3 Sparse Attention and additional LoRAs
I've tested minimax_h3_fl2v_turbo_4step_v1.3_768p_comfyui_bf16.safetensors in my existing ComfyUI workflow (of course with video flow shift of 6).
Unfortunately, v1.3 produces severely degraded/unusable results when combined with VEDA or H3 Sparse Attention, as well as additional LoRAs such as fal's Realism People and Motion Repair V2.
The important part: Replacing v1.3 with minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors or even older versions in the same workflow immediately restores excellent results.
V1.2 works flawlessly with VEDA, Sparse Attention and both additional LoRAs.
This strongly suggests a regression or compatibility issue introduced in v1.3.
Has anything changed in the distillation process, attention layers or training configuration between v1.2 and v1.3 that could explain this behavior?
Are the new weights expected to remain compatible with Sparse Attention optimizations and additional MiniMax-H3 LoRAs?
For those getting poor results with the new checkpoints, could you try increasing the Turbo LoRA strength in ComfyUI, perhaps to 2.0, and see if it helps?
For those getting poor results with the new checkpoints, could you try increasing the Turbo LoRA strength in ComfyUI, perhaps to 2.0, and see if it helps?
Good one! A LoRA strength of 2.0 for minimax_h3_fl2v_turbo_4step_v1.3_768p_comfyui_bf16.safetensors
makes the output significantly more stable. Unfortunately, the results still aren't particularly good, though...
BUT this only works without using any other LoRAs!
If I continue using h3-realism-people-t2v-i2v-v2.safetensors (strength 1.00) and Motion_Repair_V2.safetensors (strength 0.90) alongside it, I still get broken output.
This wasn't an issue with V1.2 or any earlier versions!
For those getting poor results with the new checkpoints, could you try increasing the Turbo LoRA strength in ComfyUI, perhaps to 2.0, and see if it helps?
Good one! A LoRA strength of 2.0 for minimax_h3_fl2v_turbo_4step_v1.3_768p_comfyui_bf16.safetensors
makes the output significantly more stable. Unfortunately, the results still aren't particularly good, though...BUT this only works without using any other LoRAs!
If I continue using
h3-realism-people-t2v-i2v-v2.safetensors(strength 1.00) andMotion_Repair_V2.safetensors(strength 0.90) alongside it, I still get broken output.This wasn't an issue with V1.2 or any earlier versions!
Thanks for the detailed feedback! We've noted this issue and will work to improve it in future releases.