Are YOU tired of this? 😩
Don’t have 32GB of VRAM for those big 27B models? Is your LLM prompting MiniMax H3 like it’s never seen a video model? Missing shots, no timing, dialogue nobody can tell apart, and “Music: N/A” when you asked for music?
Well, struggle NO MORE! 📺
Introducing PromptForge H3 4B Uncensored, a 4B model that prompts MMH3 like a 27B! Give it a one-line idea, or toss it a picture, and BAM 💥 you get a finished H3 prompt with:
✅ Timed shots
✅ Subject definitions
✅ Dialogue with speaker tags
✅ Sound AND music. That’s right, MUSIC!
But wait, there’s MORE! It handles all five modes:
T2VA: just type your idea!
I2VA: animate a starting picture!
FL2VA: first frame to last frame!
L2VA: lands the clip right on YOUR picture!
REF2VA: reference pictures for characters, outfits and places, every one labelled <Picture N>!
Already running MMH3 on a 12GB card? Then you’ve already got everything you need! Your 27B could NEVER! 🫵
Fine-tuned from Felldude’s Qwen3-VL-4B-Instruct-Uncensored.
🎬 Works best with darksidewalker’s MMH3 workflow
Tested in the H3 Forge, our prompt writer inside darksidewalker’s MMH3 Director node, and that’s where it shines. Point the Forge at the model, load your pictures, hit go.
Get the workflow: https://civarchive.com/models/2831978
📦 What’s in the box
PromptForge-H3-4B-Uncensored-r4.Q8_0.gguf (about 4.3 GB)
PromptForge-H3-4B-Uncensored-r4.mmproj-f16.gguf: the vision projector. You NEED it for the picture modes!
💾 How much VRAM do I need?
You do NOT need 32GB! If your card can run MMH3 (ComfyUI says 12GB cards like the RTX 3060 can), it can run this!
Measured on real hardware, with the model and projector fully on the GPU:
8k context: about 7 GB
16k context: about 8 GB
32k context: about 10 GB
Pro tip: MMH3 wants the whole card while it renders. Write your prompt, unload the LLM, then hit render. One card, both jobs!
⏱️ How fast is it?
Tested on a GPU that’s almost a decade old. Grandpa’s card! 👴
T2VA, I2VA, FL2VA, L2VA: about 4–5 seconds per prompt
REF2VA with 2–3 pictures: about 12 seconds
That’s a finished MMH3 prompt faster than you can type the idea! If Grandpa can do it, so can your card! 💪
(Times are with the model already loaded. The first run takes a few extra seconds while it loads, and bigger picture sets take longer.)
⚙️ Settings, and seriously, use these
Temperature 0.7
Presence penalty 1.5. Skip it and the model may get stuck repeating the same sentence.
Works in llama.cpp, llama-swap, LM Studio, or anything else that loads a Qwen3-VL GGUF with an mmproj.
⚠️ Content note
The model is uncensored and will write explicit scenes. All training data is adults only. It was filtered for age-coded words and young-coded settings.
🔧 Known weak spots (we’re cheesy, not dishonest)
It’s a 4B, so it can be weaker at describing what’s in a picture than the big models. Give your pictures a quick label if it gets something wrong.
Dialogue sometimes loses its speaker tag (S1). Check those lines before you render.
Squeeze a ton of action into a short clip and it can pack too many shots in.
Operators are standing by. ☎️ (Operators are not standing by. Download it.)
Credits: Felldude for the base model.
License: Apache 2.0. Fine-tuned from Felldude/Qwen3-VL-4B-Instruct-Uncensored (Apache 2.0), which is based on Qwen/Qwen3-VL-4B-Instruct (Apache 2.0).
Description
messed up the upload, when i uploaded both files they were in a single folder and civit split them, they should have been in a zip together, a rookie mistake on my end. i reuploaded the model as a zip so both files are together. also im hearing reports that the file name isnt consistent and causing issues. not sure why the files would be renaming, on my end they are the same but make sure the gguf and the mmproj file are named correctly
FAQ
Comments (27)
I ran it using the workflow you recommended and I get:
"SamplerCustomAdvanced
mat1 and mat2 shapes cannot be multiplied (152x2560 and 5120x5376) TIPS: If you have any "Load CLIP" or "*CLIP Loader" nodes in your workflow connected to this sampler node make sure the correct file(s) and type is selected."
GGUF loader by molbal (Continued Development of the GGUF Loader from City96)
The mmproj-f16 file is located next to the main one; I haven't renamed anything. The clip model just doesn't fit. What am I doing wrong?
This model is supposed to replace the Qwen3-VL-32B-H3, right? That's the whole point of saving memory, right?
@superuser111 not supposed to replace H3 clip. It's a vision model to be used for prompt writting. You can use it through LM Studio, H3 Forge, or any other LLM node. The point there is to save memory and time on prompt generation, not video generation.
what @K0DEX said
If it's not a text encoder, what is it? Is it a diffusion model? Do I replace the minimax H3 diffusion model with it? Is it a video vae? I'm trying to naviagte the Dasiwa workflow to see where I would put it but I can't figure it out. DO I still need the original minimax 3 vae, text encoder and diffusion model? Then where does this plug in?
@Light7799 Use everything else as usual (Vae, Clip, Model, etc...). Put this model into your ComfyUI folder models\llm\[FOLDER NAME with the 2 files inside]. Then, on DaSiWa's WF, you have a small bouton just above the prompt called "Prompt Forge", this is a fully automated prompt writer for Minimax H3 and is the part using this model.
@K0DEX I see, thank you for explaining that! I've been in the ai scene since sd1.4 but I haven't yet delved into video models, so this helps a lot. I've been struggling with just hardware space with everything taking up so much room, just recently got another HDD so I can download all these and give it a try.
@Light7799 Np. Yeah, space is a problem, but maybe you already know, you can put your models on a secondary disk being loaded at ComfyUI's launch. I have all my models, clips, Vaes and LoRAs on a secondary NVME to spare space on my system disk.
Is the archive missing a config.json and tokenizer files? The DSW Prompt Forge node you referenced isn't finding this.
Never mind. I think I figured it out. Needed to build the llama-cpp-python package specific to my cuda version before the backend would see GGUF files.
I dont know why but i takes 440s to make a prompt for some reason and is supposed to be fast
what im doing wrong?
440 seconds is way too long, it should be like 10-20 seconds. how are you running the model?
@TrashAI downloaded used in h3 forge in das wf model is in llm folder
maybe i need some forge setup? for it?
creativity is on 7 and inventive
for unknown reason my cpu is 100% not the gpu
@yar4ik141 My guess is it’s slow because the llama-cpp-python thats installed is the CPU-only version. You need the GPU (CUDA) version of llama-cpp-python installed into ComfyUI’s python_embeded. To build it yourself, you need Visual Studio Build Tools and the CUDA Toolkit installed. Or download LM Studio, Ollama, or llama.cpp’s server, and put its address in ComfyUI under Settings → DaSiWa → LLM servers.
@TrashAI is there a guide how to install it or a way to know exactly why that is just dont want to use lm sudio
@yar4ik141
1. Check what you have. Close ComfyUI, open a command prompt in your ComfyUI_windows_portable folder, and run:
python_embeded\python.exe -c "import torch; print('torch CUDA', torch.version.cuda); import llama_cpp; print(llama_cpp.__version__, llama_cpp.llama_supports_gpu_offload())"
False at the end = CPU-only llama-cpp-python. That’s your 440 seconds.
Note the torch CUDA number. If it’s 13.0, do step 2. If it’s 12.x, update ComfyUI portable to a current release first (it comes with CUDA 13), then come back.
2. Install the GPU version. Same folder, ComfyUI still closed:
python_embeded\python.exe -m pip install --force-reinstall --no-deps --only-binary llama-cpp-python llama-cpp-python==0.3.36 --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu130
It’s about a 230 MB download.
3. Run the step 1 command again. It should end with True now. Start ComfyUI and the prompt should take seconds instead of minutes.
@TrashAI ty soo much!
was trying to find some things but just sad later and now it works in only 14s
thx man!
and add this steps in problem solving it would be helpful for anyone else who might have the same issue
Spent hours testing it, didn’t work out well. Sticking with Grok for now.
does anyone have a workflow where this works?
where do you put it?
how do you reference it?
the dasiwa workflow had tensor names that don't seem to exist, so I can't get it to run at all. Cleaned up SIX references, fixed nodes, blah blah, still doesn't work.
Where to put it: ComfyUI/models/llm/. Make the llm folder if you don’t have one. Put the model .gguf and the mmproj .gguf side by side, ideally in their own subfolder like models/llm/PromptForge-H3-4B/. The mmproj is what lets it see your pictures. If it’s missing, the model still loads but it’s text only.
How to reference it: you don’t type a path anywhere. Restart ComfyUI and it shows up in the model dropdown on the H3 Forge / LLM nodes in the DaSiWa pack. Pick it there. The mmproj won’t appear in the list, which is normal. It gets paired up automatically because it’s in the same folder.
Requirements: local GGUF needs llama-cpp-python installed in ComfyUI’s Python, and it has to be a recent build or Qwen3-VL won’t load. The pack’s pip install -r requirements.txt covers the rest.
About the “tensor names that don’t exist” error: that happens when the GGUF goes into a model/CLIP/UNet loader, or into an older version of the nodes. It isn’t an image or video model, so don’t load it in any of the workflow’s loader nodes. It’s only for the LLM/Forge node. Update DaSiWa Nodes to the latest version through the Manager before you start repairing nodes by hand. An outdated pack would also explain having to clean up six references.
If you’d rather skip llama-cpp-python, run it in LM Studio, Ollama or a llama.cpp server and select it as a “Server:” model in the same dropdown. You set the address under Settings → DaSiWa → LLM servers.
Surprisingly it works very good with DaSiWa H3 prompt forge. Thanks for creating this model 😁
I´m confused.
Using that Dasiwa Workflow, if there is only GGUF model in llm folder, Prompt Forge says "No models found". And if there is also Qwen3 safetensor model in the llm folder, then it shows both in the Prompt Forge and I can select Qwen3 but the GGUF says (GGUF - needs llama-cpp-python installed).
But I can select a GGUF LLM model in a node called DaSiWa LLM Model Selector (Advanced).
Both GGUF's are in folder models\llm\PromptForge-H3-4B-Uncensored-r4
When I run the command .\python_embeded\python.exe -c "import torch; print('torch CUDA',torch.version.cuda); import llama_cpp; print(llama_cpp.__version__, llama_cpp.llama_supports_gpu_offload())"
I get
torch CUDA 13.0
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 16283 MiB): Device 0: NVIDIA GeForce RTX 5060 Ti, compute capability 12.0, VMM: yes, VRAM: 16283 MiB
0.3.36 True
Isn´t that saying that everything is working?
Apparently something was just changed, as simply updating DaSiWa nodes fixed it. Yesterday reinstalling ComfyUI and updating everything didn´t work but now it does.
integrated_multimodal_description: [Comment 1] The camera holds a static shot as the insane clown speaks in a rasping, weary voice (S1), <d>[English] Works fine, nice job. It's functioning perfectly for single prompts. I'm going to see how it handles my byzantine instruction set for analyzing tagged reference batches for H3. I've been trying to figure out a stable instruction set that can handle arbitrary inputs and my very specific need for chunking prompts coherently for very long runs. Even when I explicitly provide counts and overlaid captions, most models are still having trouble parsing properly. But that's probably just bad prompting on my end. I can't find the right wheel for the simple neural link node, once I get that damn thing working I can ditch this stupid keyboard too. God I hate typing.
Anyway, it does what it says on the tin. With appropriate instructions, an image and a one sentence overview expands to a dense and descriptive prompt with the proper syntax and formatting. Very quickly. </d>
overall_soundscape: Low room tone with faint electronic hum throughout. The soft whir of a computer fan and the subtle click of keyboard keys. A sharp electronic ping as the neural interface activates in [Comment 1], accompanied by a high-pitched, rhythmic electronic pulse that grows louder and faster, creating a tense, disorienting atmosphere.
non_diegetic_music: A low, pulsing synth bassline at a moderate tempo, joined by rapid, glitchy electronic percussion that builds in intensity, creating a sense of escalating digital strain.
ёлы палы
If Prompt Forge doesn't see models, here are potential causes:
• llama-cpp-python is installed as a CPU-only build.
• The llama-cpp-python build doesn't match your GPU architecture or CUDA version.
• Windows cannot find the CUDA DLLs (resulting in a llama.dll error).
In my specific case, the fix required installing a specific llama-cpp-python wheel compiled for CUDA 13.0 and the sm_89 architecture, alongside the nvidia-cuda-runtime and nvidia-cublas pip packages. This was followed by adding a sitecustomize.py script to the ComfyUI virtual environment to force the loading of the CUDA DLL paths and llama.dll on startup.