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
I'll be uploading different Minimax H3 Turbo LoRAs here and crediting the original creators
Description
FAQ
Comments (44)
Include the unpruned v4 one too because I found that it's significantly more stable - https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main
@Grockster Doesn't the unpruned need the custom node also? Or am I behind on info, I originally never uploaded them for this reason
I do have the node (if you mean the custom sampler), but I guess I'll have to play with it more to see if not having the node and having this pruned one is faster/better, etc.
How does this compare to lightx2v Minimax-h3-Turbo? specifically for I2V.
From my testing, the 8-step lora does not work with the 21gig pruned models at all. The results are very mushy awful. 4-step works fine with them but 8-step haven't been able to make anything good come out.
First of all, lightx2v_4step_turbo_768p_v1.0 is amazing model and much improved quality wise.. However, I am having issue making it work with other lora. It seems to break and starts having weird bars or effects. Anyone else run into that issue?
I haven't tried other loras but I'm having issues where the audio doesn't match the lip syncing. Anyone having issues with this?
The issue comes from LoRA training method. The model itself is guidance distilled, meaning it isn't a true model, double distillation results in model degradation using traditional training methods. We have to see what MiniMax says about LoRA training.
And someone told that wan2.2's high and low models and different loras for both - are too much complicated, but look at this now.
On todays end of the day lightx2v_4step_turbo_768p_v1.0 + h3_ref2va_pruned_bf16 + 8 steps er_sde simple works fine, at least there where other combination sucked (not tested much).
I'm honestly confused by this. Which model do I download? All of them? Just the latest published? So the one furthest to the right?
The larryvrh turbo loras a sorted by release date v4 600 ema being the latest and greatest, Then lightx2v is self explanatory v0.1-v1.0 4 or 8 steps, Personally prefer 4steps 768p v1.0
@tsolful Sorry but this wasn't helpful for me. It says the 1.0v was released Aug 11th (today) so I think this was was the latest, but I don't understand the difference between the 4step and 8 step and how are they different from the first two?
@Piraterage23 No worries, I'll try my best to explain. Larryvrh released the first Turbo LoRA for Minimax, with the first training run releasing 3 checkpoints (200 training-step checkpoint, 500, 850). They continued trying to improve the quality of the LoRA, with their latest and greatest v4 600 training steps being released. Lightx2v turbo LoRA was released a day after, with the 4-step v0.1 with their own training configuration. In the last 24 hours, they released 8-step and 4-step v1.0, the difference being what the intended steps per generation should be set. (Even though you can use the 4-step LoRA for 8 steps) The 768p is what training resolution it was set to, with the 8 step version being trained at 544p. If you still don't find this helpful, tell me which part to explain more 💚
@tsolful Just which one should I use for the best quality. ;) hahaha I really appreciate you going into detail
@Piraterage23 Hahaha i prefer v1.0 4 step 768p for realism and find myself switching between larryvrh 600ckpt ema & lightx2v v1 8step for anime
@Piraterage23 Run any turbo lora at 6-8 steps minimum, they don't have the best audio quality lower than that currently
@tsolful Are these only for FF2V? Or can I use it for Reference Images?
@Piraterage23 Im currently doing a ref2va using 8-step FL2VA v1.0 (also using fl2va model for the ref2va workflow). There is a ref2va v0.1 turbo LoRA currently, but it performs sub-par compared to the current fl2va LoRAs, which is why I haven't uploaded it here. I'll be posting the looping workflow used to create its example later today, when the run completes. It will include all the settings, prompts and references used to create the video.
I'm using the new 4-step lightx2v combined with larryvrh's ema-ckpt850-pruned from HF. Both at 0.75 strength. Seems to work well. 6 steps, simple scheduler, er_sde sampler.
Why is my audio so bad, I can't figure it out. Only the video is fine
Make sure you are updated to the latest version of comfyui, as there was a previous audio problem with low steps that was patched
@tsolful Which version? I'm on 30.0 now
this fixed my issues. Works on v0.32.0.
@ninjaware I updated it, it works now thanks guys
@felRuss Are you using different other loras as well? I can't seem to get the audio to be anything good honestly.
Audio is not working for me, the audio is always screwed. I am using comfyui 0.33.0. Without audio the generations are fine. Seems to be a general problem ?
VRAM８GBでも,時間はかかりましたが無事に10秒動画が作れました。ありがとうございます。
Apparently I'm doing something wrong.
I've tried three of the loras with 8 steps and the generation time stays the same or even increases.
I'm using the INT8 Pruned in the DAsiwa workflow.
It doesnt increase the speed. Its to make it make full quality at less steps.
Yea the turbo LoRAs are to decrease steps not increase speed, as the model is already cfg-distilled. Which was half the point of old turbo LoRAs, as models would be shipped with full cfg inference
@ForbiddenNexusAI Confused. If it takes the same time, then why not just keep the step count up??
@redlittlerabbit its same amount of time PER step. without turbo lora, say 20 steps at 1 minute per step. With Turbo lora, 4 steps at 1 minute per step. Thats where it saves time.
@4aisignups Gotcha. Thanks for clearing that up x
Anyone else having an issue with loras not working when using the turbo?
The other day I downloaded the KREA 2 int8 model from CivitAI, and it gave me subpar results. So I downloaded it again from Huggingface, it was superior to the one I downloaded from CivitAI. Are they doing it on purpose? Who knows.
if i use this lora and use 20steps.... does it has a effect? or it just stuck at 8step quality?
It gets worse. It behaves quite differently than LTX, to use another DiT as an example. With the latter, you can turn down the strength and increase steps, to a degree. With H3, 4 steps means 4 steps, at least on the boosts I've used. Er_sde/beta actually puts out a great image at full 768 resolution, model max. Converges crazy fast. Which was surprising. Save those steps for later, in case you misplace your booster.
lightx2v_8step_turbo_v1.0
It causes 5090 to run 15-20% slower making it like 5080 speed. Tried 8 steps and 12 steps. And quality of picture did not increase from other turbos that run faster.
lightx2v_4step_turbo_768p_v1.0 works, good video quality at 6 steps, but speech is often garbled. Need 10 or 12 steps to get ok audio.
Please add lightx2v_4step_turbo_768p_v1.1
Details
Files
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
Mirrors
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
minimax_h3_lightx2v_turbo_4step_v1.0.safetensors
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
minimax_h3_lightx2v_turbo_4step_v1.0.safetensors
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors