lightx2v

/

Minimax-h3-Turbo

like 614

Can you show me the video result?

This method perfectly solves the ghosting issue with excellent image quality! The quality is surprisingly good! Thank you for sharing this method

Tried this method absolutely incredible. But if you want even better quality, use the setup above with lora (lightx2v FL2V 0.1 Comfy BF16 ), 0.75 strength, and 8 steps. The quality is peak.

Try the ClownsharKsampler and change the Nois_Type_Init to "Studentt" 6 Steps ddim/beta, shift 12/10 it's incredible how good the quality is even at low resolutions. The other options from Sharksampler also bring good quality, just experiment. I use Ref 4 Step lora 0.1. I've tried a lot of things before but never really found good quality, but this is the solution for me.
Link for ClownsharKsampler https://github.com/ClownsharkBatwing/RES4LYF.git

With ETA, you can also increase or decrease the value; this also has a greater effect on quality and prompt follow. For example, I noticed a big difference when I switched from 0.50 to 0.25 or 0.75. You can generally test a lot and have significantly more opportunities to improve quality and prompt follow.

Try the different noise types yourself, they definitely make a huge difference

Try a higher or lower value for s_noise_init e.g. 0.80, that can also bring quality again or better promptly follow.

Then I tried something else that worked wonderfully and promptly reinforced it significantly and in some cases improved the quality again. There is a node for Flux2klein9b called https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer if you increase the "active scale" above, it promptly follows it much better, test it out yourself.

I tried 3.00 strength once and the prompt follow is incredible, just try 1.25-3.00 strength.
If the value is too high, the quality may suffer. But for me it varies from video to video. Just try it out. If you notice that it doesn't follow the prompt, try 1.25 first and then higher and higher until you're satisfied.
I've often had the problem before that without this node it doesn't follow it properly. With this node it really works perfectly as it should.

I tried different settings, and a value of -0.10 makes the movements more realistic for me.

Try the other values too, it works almost everything and affects the quality or prompt follow.

I haven't noticed any change in these controls so far; everything else works great.

There are other nodes available in the custom node, but I haven't tried them yet.

I've only tested it with the reference model so far, but I think it will work great with the other model as well. Feel free to write to me and ask if it works for you and if you could improve your quality or prompt follow, I would be very interested.

Thx for sharing! The clownsharksampler with opt alone generates quality beyond expectation!!

Can u send me the workflow please?.

@SatoruGojo1
I don't need to send you a workflow, you simply replace the Ksampler or Customsampleradvance with the Clownsharksampler, then you simply add to the options what I showed in detail in the photos, that's all it is :) it's actually very easy.

video demo ?

hello, is this compatible with 20 steps, bf16 full model and no any speed loras and nodes ?

@sananeaqr I haven't tested it, but normally it should work great at 20 steps without turbo.

Thank you, it works great for the video, but if you use an audio reference, the sound is very poor...

@jovanny47 Try changing the shift in audio. I had that before and managed to get it working again with shift. I also have to say that I haven't really tested everything yet, but I found the quality to be much better; we'll have to see how we can make everything even more perfect.

capitan01R/ComfyUI-Flux2Klein-Enhancer

placebo from a developer not understanding what they have actually been doing, and promoting certain harmful and redundant practices

Feel free to write to me and ask if it works for you and if you could improve your quality or prompt follow, I would be very interested.

I am in, could you continue investigating such things so we are sure what could really enhance something? I just use models as they are without any custom nodes (excluding something like negpip because it allows me to get noticeably better outputs when, for example, a model puts noise, watermarks, motion blur, etc. because when CFG is 1 or close to it, or a model doesn't support negative prompting, it helps enforcing it)

@dummy9996 I am testing daily and if I find improvements I will write them here :)

You indeed found some great settings. I only put the ClownsharKsampler in the standard workflow with the settings as you mentioned and the generated video is a lot better. Less malformed limbs when movement is fast the overall quality is just better. Used it with the 8 step v1.0 lora at 10 steps.

@hanswurstwerner Thanks for the tip, It seems to be working on shift 4–4.10, the sound on the background seems to bad..need to test it. i use minimax_h3_hybrid_fl2va_ref2va_b20-49-int8 and minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy - 0.75 st

here's how it is connected:

What I am writing here now is with the fl2va txt2video

I've tested the other nodes and will definitely try this one in the photo. Make a short video that generates quickly and then try different values. I tried a very short prompt and the quality and prompt follow are mega good.

This one is really great too. It's enough to increase the value at the top by just 0.20 to make it follow the prompt better. I tried it without it, and instead of running, the person went for a walk. Then I increased the value by 0.20, and the person runs, just like I wrote in the prompt.

A small increase of, for example, 0.10 in contrast gave me significantly more sharpness in the image.

What I just tested, leaving everything on normal and just increasing the contrast, it's absolutely insane how well it follows the prompt now, for whatever reason I can't explain it, the image is ultra sharp and the prompt was implemented exactly as I wrote it.

And I finally tried this and it's just incredible, try it yourself.

These are my current settings after a lot of trial and error for txt2video, it's just fun the videos look so realistic and good.
Turbo Lora 0.80-1.00 8-10 Steps

What I am writing here now is with the fl2va txt2video

I've tested the other nodes and will definitely try this one in the photo. Make a short video that generates quickly and then try different values. I tried a very short prompt and the quality and prompt follow are mega good.

This one is really great too. It's enough to increase the value at the top by just 0.20 to make it follow the prompt better. I tried it without it, and instead of running, the person went for a walk. Then I increased the value by 0.20, and the person runs, just like I wrote in the prompt.

A small increase of, for example, 0.10 in contrast gave me significantly more sharpness in the image.

What I just tested, leaving everything on normal and just increasing the contrast, it's absolutely insane how well it follows the prompt now, for whatever reason I can't explain it, the image is ultra sharp and the prompt was implemented exactly as I wrote it.

And I finally tried this and it's just incredible, try it yourself.

These are my current settings after a lot of trial and error for txt2video, it's just fun the videos look so realistic and good.
Turbo Lora 0.80-1.00 8-10 Steps

Awesome, but still I get slow motion 😞

how does it work with fl model?
edit: Motion is better, but idk about the quality, 0.5M looks close, i can't really tell

@rkss @Jam747

I took a screenshot of my workflow for you, you can try it that way. I only have the slow motion thing when the prompts are too short. I usually have Ki write the prompts and just edit them so that you have a template. For example, I tell Grok to read Minimaxh3's prompting guide and then write me a prompt with a shot and timeline for 5 seconds or 10 seconds. This works perfectly, here is a template:

integrated_multimodal_description:

[Shot 1] Live-action, [STYLE], [CAMERA / FRAMING]. [Who / what is in the scene]. [Environment]. [Action starting now]. [Camera movement: type + intensity + speed].
[Shot 2] At 00:02.000, the camera cuts to [new framing]. [What changes / next action].
[Shot 3] At 00:04.000, the shot continues / cuts to [final action or ending].

overall_soundscape: [Ambience], [physical action sounds], [breathing / crowd / wind / impacts if needed].

non_diegetic_music: [Describe music] / N/A

https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler/tree/main

any one tried this?

I got some REALLY good results, starting at 0.3mp and do second pass to 1.2, faces from distant looks decent.

creating A/B/C workflow, testing this clownshark method, testing the latent upscaler method and combination of both.

so total of 3 gens, not an expert so my settings wont be as correct as i think, will publish workflow as soon as i have it ready on my github repo or civitai, will edit this comment when i have it ready.

if someone managed to test it, i would love feedback maybe could tweak the settings to get even better results.

@hanswurstwerner

Beta57 gives me the best quality result, but prompt adherence is off. model still does what is asked, but differently