lightx2v_8step_turbo_v1.0+ lightx2v_4step_turbo_768p_v1.0 uploaded; works with pruned/non-pruned
Current best settings: 5-8 steps er_sde/res_multistep simple. Strength 1.0 seems to be the best with v1.0 for v0.1: 0.75 (0.5-0.8)
New larryvrh v4 600_ema uploaded.
This LoRA is an early version of a training run currently underway that larryvrh is doing. Degrades audio quality at 4 steps.
Best used at 6-8 steps res_multistep + simple
Examples are generated with 1.0
I did not train, distill or create the original Turbo LoRA weights. This model page only provides modified compatibility versions intended to allow the remaining compatible LoRA adapters to 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 lightx2v team and kijai for the comfy compatible version
I'll be uploading different Minimax H3 Turbo LoRAs here and crediting the original creators
Description
FAQ
Comments (32)
The older "h3_turbo_4step" is a copy-paste from drbaph, but what exactly is the new lightx2v_turbo_4step_v0.1 (1.82GB) and where does it come from?
I upload to CivitAI as I would only use CivitAI to get models in the past, and some people might still do. I do train loras here and there and int4mixed quantizations myself for low vram users + resources on posts
@tsolful Thanks! So the Kijai version should be the gold standard. I need to test the best combinations of model+steps+strength+sampler, now I have no idea what would work best. 4 steps worked fine for me yesterday with drbaph's LoRa (plus the custom node to fix the sound), but another day another LoRa, so there is no end to it...
@bnzarev821 Kijai pushed a update to fix audio problems to the latest comfyui, but you would have to update your comfyui with " git pull origin master "as it is still in the nightly version of comfyui. Ive found 8 steps, strength 1, with ckpt500 of larryvrh's version and 6-8 steps, 0.75 with the new lightx2v lora to work well with res_multistep simple (it is best to keep the scheduler at simple in my tests but you can test other samplers)
@tsolful Some recommendations from the Kijai's model card:
4 steps + 0.75 strength + er_sde or sa_solver
Update: just tested with a fixed seed and er_sde looks more natural than res_multistep, at least for 4 steps + 0.75 strength, but it's too early to generalize I guess. Also er_sde seems a little better than sa_solver for this patricular seed, but both are very close. Euler is also good, so there's a lot of choice.
@bnzarev821 Been using er_sde for "realistic" outputs and res_multistep for anime, sa_solver is hit or miss for me
Great and nice examples💕👍🏻
lightx2v_4step_turbo_v0.1+ rank21 and 850ckpt. are good. but if you use them to expand a video, there will be a huge contrast addition between original video and expanded scene. 500ckpt, keep contrast unchanged
Appreciate your results, hopefully with the updates it gets fixed
Cant get any of it work, it always produces utter garbage :/ tested with different models and different turbo loras its always utter mess
maybe just disable sage attention, easy cache, spectrum nodes if it is enabled
also remove the sage attention cmd line from run.bat file
8 Steps minimum with strength 1 chkpt500
v4_step600 without ema version?
uploaded 💚
newb question what does ema do or dont do ?
@jynkz41 It is a training technique that tracks a smoothed out average of the model's weight values over time rather than just the pure fluctuation which is the non ema
@tsolful ok so is one effectively better than the other then ?
@jynkz41 The actual author recommends the ema
@jynkz41 in theory, the early checkpoints, EMA is better, vice versa as training continues. Best to test both outputs and decide really. as @pijseh2244955fdfd said the author recommends v4 600ckpt ema for 6-8 steps and for the specific case of 4 steps and heavy motion, the older v1 850 checkpoint can still be the friendlier pick. Personally have been using Lightx2v Turbo LoRA as audio fidelity is better imo
I stress tested this hard at 1.0 MP for 13 seconds - 503.13 secs later had a beautiful render - VRAM held at 98% consistently the entire way - running a 5090 32GB card with 96GB ram. I will post the result in a moment.
Nice with sage attention that about how much it takes for .9 at 720 for 10 secs does the audio hold up as well and are using sage in the generation ?
@Hzoid I did use Sage attention. There is a guide on hugging face I think that recommends some adjustments for both the audio and video drifts. Let me dig them up so I don’t recommend the wrong thing and I’ll DM them to you. Without that you will have garbled audio for sure
I did multiple videos with I2V and it works great but is this only for FL2V model or is it compatible with REF2V also?
i'm using it for Ref2VA
no issues
using Chkpt600 Lora
8 Steps 1MP
from Native 15FPS to 30FPS interpolation+RTX UPScale took around 30minutes for 10sec clip
bcoz of interpolation fast movement's are not good
but slow motion face closeup shots are awesome, bcoz of 1MP
Config-48GB System RAM RTX 3060 12GB
AI Model Files are stored in 7200RPM HDD
OS and Paging File Set to Nvme SSD
Awesome! But how to increase Audio-Quality? Using 8 steps, lora-strength 1.0
Update comfyui
Does this work with the refrence photo imput to output model as well? the one that does multi photo inputs? any fast answer please and thank you.
Yes both of them work, ill upload my first test using lightx2v turbo on reference model and link here
Massive difference, almost half the time of generation with very good quality, thank you very much. Using the Lora at 0.75 with 8 steps.