We're a Day-0 launch partner for MiniMax H3, a next-gen open-weights multimodal video model - and it's already in the generator. No waitlist. Open weights land in a few days.

MiniMax H3 is live: open-weights video, generate now

Work in progressThe models with 📦 mean that are archived and i have yet to updateCheck the About this Version of the chose workflow for proper introductionquestion, how do i embed the nodes into the images or videos?if there is a way to load OmniGen/CogVideoX/LLM/t2a/AnimateDiff(gpu) thru a CustomSamplerAdvanced let me know please

I generate the input image with a pdxl model but you can use your favorite t2i, i use v-prediction for maximize creativity, i'm using my favorite noise chain for t2i thats 1 step fe_heun3, 1 step SamplerSonarDPMPPSDE(student-t), 2 steps lcm (uniform), for better first step you can use the SamplerDPMAdaptative node thats left alone, it's optimized to go fast, but you can play with it, for the second step you can prolonge the lcm(uniform) for more smooth results but less creative or add a SamplerRES_Momentumized(highress-pyramid) and finish with 2 steps of lcm(uniform), you can also try the ClownSampler node for step 2 to get a different result, the lcm(uniform) can also be changed for the ClownSampler but i really like what lcm(uniform) does. Now for the ltx video, you don't require the sampler chain, but if you want the best from the model, experimenting is your best bet, also, the cfg modulates the movement, the consistency and the artifacts, you may as well experiment with different cfgs for each 1/3 of the generation, thats also a reason for the split sigmas, that improve the generation by a lot

🦄🖼>ltx📹

Experimental MochiMochi is native on the latests ComfyUI versions, git pull or update.Sampler:I recommend chaining samplers of different lines to get the best of each per step, the order is important, i tend to start by 1 step of dpm_adaptative for the success rate, 1 step of sonardpmpp (student-t) for creativity and finish with 3 steps of ipndm &amp; 4 of ddim.In my examples the Sampler chain is 1 step of dpm_adaptive &gt; 1 step sonar_dpmpp_sde (student-t) &gt; 1 step lcm &gt; 2 steps SamplerDynamicCFGpp &gt; 4 steps heunThe sampler chain for the 5gen/3d style vid is 2 steps SamplerHeunCFGPP &gt; 3 steps idpm &gt; 4 steps ddimdpm_adaptive|fe_heun3|ae_dopri5(+Guide)sonar_dpmpp_sde (student/pink|power/uniform)|dpmpp_dual_sde_momentumized| dpm_2_ancestral(+Color)SamplerDPMCFGpp|heun_cfg_pp|distance_s4e1_cfg_pp|SamplerCCCFGpp/SamplerDynamicCFGpp|SamplerHeunCFGPP|uni_pclcmdpmpp_2m_sde/*_3m_*|*_gputcd_wipndm/majority|heunSamplerX0CFGppto speed up, I normaly generate at 5 steps, and when I find something good I generate it again with at least 9 stepsNoise-types:3D noises (pyramids,perlin,fractal,voronoi,etc) or laplacian dont workWork: uniform,power,pink,studenttBad: green_test, gaussianVAE:You may get an Out of Memory, Until a tiny auto-encoder launch for genmo mochi, Prevent with the argument "--cpu-vae" but takes 6x longer, or use the "latent upscale by" for a fast output (expect blur), or Save latent and decode later.CFG:For a successful generation i suggest a cfg greater than 1,You can get a good image with cfg 1 with a good sampler chainScheduler:Use a primitive for randomizing the alpha and beta parameters for the "BetaSamplingScheduler" until you find your favorite.Specs:I have a NVIDIA GeForce GTX 1660 SUP, 6GB vram (like 10 times slower than a 30**/40** series), if i can use this you can as well, probably if you have a good graphics and vram, you may use the correct resolution, correct length and VAEDecode, i didn't try it so it requires experimenting, if you have 4GB vram you can try reducing the length and resolution, but i didn't get good results when testing.

📦🍡

FluxS🦈

🦈🎭🦄

Lora 🛍️

Sadly not compatible with CustomSamplerAdvanced so i can't use noise types nor sampler chains nor custom scheduler, please let me know if there is a way of using it, also this is considered Old as <a target="_blank" rel="ugc" href="https://huggingface.co/THUDM/CogVideoX1.5-5B-SAT">CogVideoX1.5-5B-SAT </a>launched, rooting to try it.

📦Cog🎥X

This workflow is messy, you can select to generate the first steps with sd35 or flux(1or2or3 steps) and end with pdxl(2steps), the models use the noise types and samplers that in the moment i thought were better, also you can make the pdxl hotshotxl for I2V (WIP)

🎨

It can make images at 1 step, and is also good with more steps, i started with a working 512 x 512 but i don't know if it still works at that

🏇

Stable Audio Open ExperimentalWIP (I didn't try enought)Samplers:Res_Momentumized (Creative)dpmpp_sde_gpu (best)ddim/others (wip)Sampler chain don't work as well as customsampleradvanced chains for this modelNoise Types:green_test, uniform (wip)

📦S🔊🚪

At model unload it losses the seed (noise type chain fault), Vector Sculptor don't work at fp8_e4m3fn-text-enc, i didn't test the model at fp16, experimental, if you want to quantize your checkpoint first separate your unet with ModelSave node, then run .../ComfyUI/custom_nodes/ComfyUI-GGUF/tools/convert.py --src 'your model', finally .../llama.cpp/build/bin/llama-quantize 'your model' Q4_K_M (or other).

📦🏔️

Adds this cmyk style to the output, At model unload it losses the seed (noise type chain fault), Vector Sculptor don't work at fp8_e4m3fn-text-enc, i didn't test the model at fp16, experimental, if you want to quantize your checkpoint first separate your unet with ModelSave node, then run .../ComfyUI/custom_nodes/ComfyUI-GGUF/tools/convert.py --src 'your model', finally .../llama.cpp/build/bin/llama-quantize 'your model' Q4_K_M (or other).

📦💫

📦FluxS

you may want to remove the score tags from the Neg PiP to get color.It requires <a target="_blank" rel="ugc" href="https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb">https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb</a> for the NegPiP and <a target="_blank" rel="ugc" href="https://github.com/WASasquatch/FreeU_Advanced">https://github.com/WASasquatch/FreeU_Advanced</a> for the FreeU_Extreme, the freeU extreme is the culprit of the motion that's happening.

📦🎭🦄

4 or 9 steps, Sampler chains & noise types chains - Flux, Mochi, PDXL, MoGe, LTXV, Others