It's uncensored, but it takes a while to make the videos, although it works well with low VRAM on lightweight models; we need a Lora Turbo.

@tsolful what? u sure

@tsolful its definitely has censor in it, i tried img to video, it make pussy looks weird, idk how in the galery the make it looks good

@Seii1 yeah u need to be descreptive of the thing he does have to guess when he guess something he does it shitty but if tell him this is like this and this this way it make a wonderful work ask gemini for help and tell him what u doing he tought me it (i still too lazy to do it right but)

Yes, it's everything you say here I have been using since yesterday

https://civitai.red/posts/30163009

私のところではあまり良くないですね…画像に映っていればうまくいくのかな。

Yes! This model—with an 8-step LoRA, realism-enhancing LoRAs, and NSFW LoRAs—is going to be insanely amazing!!

@tsolful nice. whats the performance?

Works fine even on my rtx 3060 12gb +16gb RAM. You only need a fast NVMe SSD and to manually set a big (150-200GB) virtual memory file (pagefile.sys). My test here

@tsolful can u share the worflow and teh GGUF's what u have used in it

@bnzarev821 can u share the workflow? we have same rig

@tsolful Thanks a lot! By the way, how much is the generation time/ per 1 output?

as a matter of fact im running it on a rtx 3060 12gb 32 ram

@tsolful 5060ti16g+32g有压力吗

ありがとう！同じ見解でよかったよ。これから有志で多くの人がLoRAを開発してくれると思います！待ちましょう。

@Japamelon 作ったらお互い配布しましょう！！

@tsolful do you know if a vj and pen lora can be made with just images?

A competitor? The Minimax H3 simply destroys the LTX 2.3 in every aspect (general anatomy, complex movements, visual consistency, physics, sound), plus it's omni... Once we get 8-stage LoRAs and others like enhanced realism LoRAs and NSFW LoRAs, it's going to be absolutely fantastic!

More than that, I'd say... I was never fully sold on LTX2.3, but now with H3, I might forget about it entirely. I'll just be waiting for community support and refinements and then make a decision.

Does this mean I should propably say goodbye to Wan 2.2? Or how is it? (I know Wan does not have sound)

@HugMeIntoFace Not yet, for sure.... give this a month or so so the community can do merges and loras, and then re-evluate.. I love Wan2.2, but yeah, the lack of sound has been its main drawback this entire time.

@HugMeIntoFace Absolutely! You can go ahead and delete Wan and LTX right now! Just look at what Minimax H3 does without any tweaking (no 8-step LoRAs, no concept/NSFW LoRAs, and no custom workflows)—the results are already WAY BETTER than LTX and Wan combined! Just imagine what it’ll be like with LoRAs and custom workflows... It’s going to be FANTASTIC! :D

@HugMeIntoFace Go to your LTX and WAN folders and just fucking delete. You won't go back.

@ArttTaku A month!? If you can't tell within the first video you make... LOL

@AI_2_addicted until now it is just slower, and I need around 50GB Ram to render some vids at 20 steps... lets hope the best. :)

@qingyantrong4932325555 @AI_2_addicted Guys I am too old to do impulsive f***s that I could pity later. But I love your enthusiastic aproach. I will look forward what MaxH3 could deriver, in the mean time I will keep the Wan 2.2 warm, it would be heartbreaking moment, if I lost it. I hope that this NEW stuff can continue what I already learn. (I have high demand for consistency for characters and I almost perfected it in Wan 2.2, anything else would be a downgrade)

@tsolful What does: "But were Finetuned/Trained further for higher quality outputs on the intended Use" mean?

File SHA tells me that files that You uploaded are EXACTLY the same as ones in comfyUi.

So You DL files from comfy and shared them? Is that correct?

@tsolful put all of the acronyms in the description please. FL2VA REF2VA you explained here.. but R2VA? That's different than REF2VA? What's VR2VA? Virtual Reality to Video & Audio?

If I only have 1 image, can I use FL2VA to generate?

@tsolful I use the minimaxH3INT8_fl2vaINT8Pruned model, and only 1 input image. Am I use the wrong model?

If I choose the model, and only have 1 image connect first frame(no last_frame image),can it work? or I need download the REF2VA MODEL

I use the comfyui template workflow,But already the same error:


[ERROR] !!! Exception during processing !!! The size of tensor a (59940) must match the size of tensor b (1451808) at non-singleton dimension 2

[ERROR] Traceback (most recent call last):

File "C:\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute

output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)

File "C:\ComfyUI_windows_portable\ComfyUI\comfy\model_sampling.py", line 97, in noise_scaling

return sigma (s noise) + (1.0 - sigma) * latent_image

~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

RuntimeError: The size of tensor a (59940) must match the size of tensor b (1451808) at non-singleton dimension 2

I solve the problem,ComfyUI_smZNodes or ComfyUI-TiledDiffusion have some trouble with it.

@sekaiwlc07860 I ran into the same problem; how did you solve it? Thanks.

why do you want mxfp8 when int8 convrot is better quality and gives you the same speed?

@MrFlex If you have a 50 series card mxfp8 is faster than int8 convrot

@jervis314 i have a 50 series card and its not faster, and way worse quality, nvfp4 is faster but really bad quality

@MrFlex I have a 5090 and mxfp8 was definitely faster than int8 when I tested it with LTX and Krea 2. There was no difference in quality.

Lets hope so...

Toolkit already supports it: https://github.com/ostris/ai-toolkit/commit/8502a845b16418b1042dd7e3b2d819cc86326b1c

4060 8gb run ?

For which one of these 4 models are you referencing?

@BinaryBottleBake REF2VA INT8 pruned.

@Samoht yea

Strange, i use a 3060 12GB card and my gens were alot faster using the default workflow provided by comfy. I have since reworked it to use EasyCache and Sage and its faster, but 30 mins for a 10 second gen on a 3060 doesn't sound right.

I was generating at 0.7MP 16:9 Widescreen FirstFrame with default workflow at around 8-10 mins.

Are you using Int8/int4 or FP8?? An RTX 3060 12GB shouldn't take 30 minutes to generate a 0.4 MP image 10 seconds, int8.

Are you using atleast CUDA 13.0 (cu130) in your ComfyUI? Also, is the Comfy backend CUDA or Triton enabled , (disabled = false)? you can see Comfy backend info at the comfyui startup in your terminal.

If so, you should be getting the INT8/INT4 speedup. A 30 minute generation time on 12GB vram suggests that the CUDA or Triton backend isn't enabled (disabled = true), so the optimized kernels aren't being used.

Indeed. I dont think i want to return to LTX now. I have so much fun playing with this one.

Goonatron 6000 Pro =)

It actually can do porn without any loras

I have deleted all models of LTX2.3 from my pc