Krea 2 + MiniMax H3 AIO
V1 is faster than V2 but less quality.
If you’re good at prompting and you just want one button, this is that workflow.
You type the still into Krea. That image feeds straight into MiniMax H3 as the first frame. H3 does the motion and the audio, then a 2x latent refine, then FILM up to 48 fps. One queue. You’re done.
Krea box is the picture. H3 box is what happens after that picture — one action, camera, sound. Don’t describe a different opening shot than the one Krea just made.
Match the two aspect widgets. Leave last frame alone unless you actually have an end pose. Duration is in seconds at 24 fps internally, then it gets interpolated to 48.
H3 megapixels sit at 0.2 on purpose. That’s the cheap draft. The upscaler is what makes it big. Don’t crank that number unless you’ve got VRAM to burn.
Models: museByStableYogi_v35Int8Extended.safetensors, qwen_image_vae.safetensors, qwen3-vl-4b-heretic_nvfp4.safetensors, mmh3-4b-ClipProj-celeb-mlp.safetensors, 10Eros_Max_h3_TURBO-hybrid_beta5_int8.safetensors, minimax_h3_video_vae_int8_convrot.safetensors, minimax_h3_audio_vae_fp32.safetensors, minimax_h3_latent_upscaler_3d_bf16.safetensors, taeh3.safetensors, film_net_fp16.safetensors
The H3 encoder has to be Qwen3-VL, not plain Qwen3.
Nodes: native MiniMax H3, KJNodes, ComfyUI-ClipProj, Sol-Attn H3, PlagueKind LTX/MiniMax LoRA loader, ND Super LoRA Loader, LTXVideo, MiniMax H3 Latent Upscaler 3D, FILM interpolation, ResolutionSelector, ManualSigmas
Krea and H3 LoRA slots are already on the canvas.
The first frame is a guide, not a paste. Prompt the motion from the still. Pass 1 is rough on purpose. Pass 2 is where it cleans up. The preview can look a little softer than the file you save — trust the mp4.
Uncensored stack. Don’t be an idiot with it.
Description
this workflow combines the krea 2 generation workflow into the H3 generation workflow to create an all in one generation hub for going straight from a prompt to an image to a video in one generation
gonna need some nodes to get it running, and if you have issues leave a comment and ill get back to you ASAP
custom nodes needed(as far as i know):
KJNodes
RTXSuperResolution
rgthree-comfy
interpolation model:
film_net_fp16
maybe some minimax nodes, but once the workflow is open, it should show you the errors tab which you can copy the nodes needed into google and find them. otherwise, feel free to comment.
FAQ
Comments (11)
i am uploading a better version of this soon
great idea. worth testing.
there seems to be an issue with the second pass's first few frames, but i've fixed it on my other workflow and im going to upload both fixes today.
@dankyouverygud Ok. Thanks. Couldn't test v1 yet. will try v2 asap
@denolim465778 sounds good, let me know how it works for you!
@dankyouverygud just got errors, first clip proj, then customspampleadvanced or something. i gave up
@denolim465778 which workflow?
@denolim465778 you need this in the clip projector on the load clip node: mmh3-4b-ClipProj-celeb-mlp.safetensors
@dankyouverygud thank you. It was the AIO workflow, i have tried mmh3-4b-ClipProj-mlp.safetensors. mmh3-8b-ClipProj-mlp.safetensors and the 3.1 version of them, no one worked. i will try with with the one you recommended again.
is there no chance to work with those i have downloaded already like the 8b version?
and other diffusion models or checkpoints?
@denolim465778 the clip proj is specific to the clip model type, the 4b works projecting into the 32b but not the 8b.
@denolim465778 the clip from qwen needed is qwen3-vl-4b. whichever you prefer to use