minimax_h3_dmad_full_dareties_v4_step600_comfy experimental merged myself, of the current best loras uploaded.
Current best settings: 6-8 steps, er_sde/lcm, simple, Strength 1.0 Video/Audio Shift 12/3
dmad_minimax_h3_4step_full_critic_comfyui uploaded (On par with larryvrhv4 video but better audio quality imo)
Current best settings: 6-8 steps, er_sde/euler, simple, Strength 1.0 Video/Audio Shift 12/2
taomate_fl2va_3step_ema uploaded
Current best settings: 4-6 steps, er_sde/euler, simple, Strength 0.75
minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16:
Current best settings: 8-12 steps, er_sde/res_multistep, simple, Strength 1.0
lightx2v_4step_768p_v1.1+ v1.2:
Current best settings: 5-8 steps, er_sde/res_multistep, simple. Strength 1.0
larryvrh_v4_600_ema:
Current best settings: 6-8 steps, er_sde/res_multistep, simple. Strength 1.0
All examples are generated with 1.0 strength
I did not train, distill, or create the original Turbo LoRA weights. This model page provides only modified compatibility versions intended to let the remaining compatible LoRA adapters load with ComfyUI's built-in LoRA loader when using the pruned/curve-form model.
Full credit for the original h3_turbo_4step_pruned belongs to larryvrh. And drbaph for the pruned comfy-compatible version
Full credit for the original lightx2v_4step_turbo_v0.1 belongs to the lightx2v team and kijai for the comfy-compatible version
Full credit for the original taomate_fl2va_3step_ema belongs to the TaoLiveAIGC team. And drbaph for the comfy-compatible version
Full credit for the original dmad_minimax_h3_4step_full_critic belongs to Bytedance & Texas A&M University students.
I'll be uploading different Minimax H3 Turbo LoRAs here and crediting the original creators
Description
FAQ
Comments (12)
Tested the lightx2v_4step_768p_v1.0 today.
Previous turbo was the minimax_h3_fl2v_lightx2v_v0.1_dareties_v4_step600
previous stats: 10sec 0.6mp @ 576x1024 8steps = 289secs render time.
current stats: 10sec 0.6mp @ 576x1024 4steps = 177secs render time.
both tests ran identical settings outside of steps.
38% faster - with no real quality loss.
running 32gb ram with 5060ti 16gb
Could you please specify the exact production parameters? Except for the number of steps.
I also ran some tests compared to minimax_h3_fl2v_lightx2v_v0.1_dareties_v4_step600.
At 6 steps. This Turbo LoRA was slightly faster. But the quality drop was very noticeable. I am sticking with minimax_h3_fl2v_lightx2v_v0.1_dareties_v4_step600, with better quality as a very slight time increase.
@archerjones1985 That is interesting you lost that quality!
What card are you running?
Kinda wild how much variance there is.
Another thing I noticed was that with the dareties version I was running AdaLN patching and it would typically attach ~300 patches - when I ran this new version, it only attached 208 patches and had no mismatched patches at all. So it is considerably cleaner.
@koalabearr hey mate - all of my workflows use sparse attention set at 0.95, I use shift ideo at 8.0, I use chunkfeedforward at 2 chunks & 4096 threshold & SLA attention. Another part of it is the Vae/text encoder combination. I have an INT8 video vae and use an nvfp4 qwen text encoder. I will swap between samplers and schedulers depending on results - usually euler/simple, er_sde/simple or res_multistep.
Then there is also reference image size relative to the output resolution - if you are rendering in say 9:16 - make sure your input image matches, it just makes it smoother and faster.
My results are without anysort of upscaling like FSR or interpolation - If I get a render I like I will lock the seed and re-render with it turned on, sometimes adds a minute or more.
light2x v1.1 768 got much worse in quality than 1.0 which is the best so far (IMO) with same workflow.
7steps. tried all recommended schedulers
Yeah, I'm finding 1.1 is worse than 1.0 also
Thanks for sharing these regardless of results, I will test
I did a quick comparison with generated videos of a girl shaking her tits.
On the model minimax_h3_ref2va_pruned_int8_convrot.
In general, I compared loras:
larryvrh - minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors
larryvrh - minimax_h3_turbo_4step_ema_ckpt850_pruned_comfyui.safetensors
lightx2v - minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors
lightx2v - minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
lightx2v - minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors
lightx2v- minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
shift video\audio = 6\3
euler\simple
4 steps, for testing
larryvrh step600 - str=1. Generates 5 seconds more or less normally. If I set it to 10 seconds, I get artifacts and "vibrating" animations instead of normal movement.
larryvrh ckpt850 - str=1 - 10 seconds of video - artifacts - the camera jumps along with the boobs. I increased shift to 12 - there are fewer artifacts, the camera doesn't jump, but there is still some flickering in the background, and the boobs shake slightly. I lowered str to 0.75 - the camera jumps again, the boobs shake normally.
Shift 6 + 0.75 str. lora = earthquake effect again.
lightx2v fl2v 4steps v0.1 - str=0.75 - the boobs don't shake. Okay, I didn't continue testing.
lightx2v fl2v v1.0 and v1.1 - on str=1, the boobs barely jiggle. On str=0.75 the movements are normal, active, and the boobs jiggle (I didn't notice any visual difference between v1.0 and v1.1).
lightx2v ref2v v0.1 - str=1, the boobs don't jiggle. On str=0.75, the boobs jiggle normally.
It's hard to judge the quality visually; more testing is needed.
I also tested larryvrh loras with the recommended settings:
res_multistep\simple
6 steps
larryvrh step600
video shift = 6, str 1.0 - slightly overdone, the lights from the disco ball are rotating in the background, which isn't in the prompt. The boobs barely move.
video shift = 6, str .75 - the light from the disco ball is gone, now the confitti are falling from the ceiling, but the movement is still minimal.
shift = 12, str .75. The confitti are gone, the image is normal, but the movement is minimal.
larryvrh ckpt850
shift = 12, str .75. It looks normal, but the boobs don't move at all. I didn't test it further.
Which is why I only recommend using the full 1.82GB model and not the pruned 500mb version. The quality loss is huge and it's not any faster. Also, I usually run 0.1 and 1.1 at 0.25-0.35 and basically everything looks much better and natural for me. I do however run higher steps because I don't think 6-8 are enough for any thing. Higher steps tends to result in better prompt adherence across the board.
Running Turbo loras at full strength isn't recommended anyhow. You want to run as low as you can to match the steps you're trying to run with the best quality possible before it starts to ruin physics, aesthetics, and prompt adherence.
I'd start low and move up. 0.25 is a good spot, and move up 0.05 until you notice it looking as you like.
@DaddyWolfgang There can be many extremes. With one lora, movement is generated normally, but fluids are washed out, simply "erased" with each step. With another lora, fluids are generated more or less, but movement is sluggish. This isn't specific to these loras, but a trend I've noticed before. For the WAN 14B , I only saw a cool turbo lora that generated simply awesome liquids, especially if raise the cfg to 1.5-2, could get cummogeddon.