DJ V Prediction Param

You load NoobAI vpred - or any Illustrious-family vpred finetune - as a GGUF, hit Queue, and get flat gray junk instead of an image. No error, no warning, just wrong. That's the signature of a v-prediction model being sampled as if it were epsilon, and it's exactly the problem this node exists to solve.

Here's the thing: normal SDXL vpred checkpoints tell ComfyUI who they are. The weights carry a v_pred key, and ComfyUI's loader reads it and switches to V_PREDICTION sampling automatically (a ztsnr key also flips on zero-terminal-SNR). But when you extract the unet and quantize it to GGUF, that metadata doesn't survive the trip. The model loads, ComfyUI assumes epsilon, and a vpred model run in epsilon mode doesn't fail loudly - it produces static. The author of this pack hit it while quantizing NoobAI vpred to run on a 3GB potato GPU, and this node is his two-click answer.

What it actually does

DJ V Prediction Param is a model passthrough with a dropdown. Two inputs:

• model - the MODEL output from whatever loaded your unet, GGUF or not.
• parameterization - epsilon or v_prediction.

One output, MODEL, goes straight into your sampler.

Under the hood it clones your model, sets the internal parameterization to "v", and patches in a ModelSamplingDiscrete + V_PREDICTION sampling combo. The clever little bit: there's no branch for epsilon at all. Pick epsilon and the model passes through untouched - so the node is effectively a toggle, and it only does anything when you flip it to v_prediction.

When you need it - and when you don't

Don't slap this on a normal .safetensors vpred checkpoint; ComfyUI already handles those automatically and you'll be adding nothing. Reach for it when the metadata is gone: a GGUF-quantized unet (from the author's SDXL GGUF Quantize Tool or city96-style quantization), or any custom-extracted model whose state dict lost its v_pred key. If your quantized vpred output comes out looking like a gray blob, this is the first thing to try before you blame the quant.

Installing it

ComfyUI Manager, or the plain clone:

cd ComfyUI/custom_nodes
git clone https://github.com/magekinnarus/ComfyUI-V-Prediction-Node

Restart ComfyUI. That's the whole install - there's no requirements.txt and no pip dependencies; the node only uses ComfyUI's own internals. One note: the README's clone line points at the older ComfyUI-DJ_nodes.git family repo, so if that clone fails, use the URL above. And this pack has no GGUF loader of its own - you still need ComfyUI's built-in GGUF support or city96's ComfyUI-GGUF to load the quant in the first place.

Where people get burned

The node fixes the parameterization, not the CFG. vpred burns at the CFG values epsilon models love, so if colors are blowing out or the image is cooked, drop CFG to roughly 1.5–2.0 with a CFG++ sampler, or add a RescaleCFG node around 0.7. This node doesn't touch zsnr, and it won't save you from stacking non-vpred LoRAs on a vpred base. And honestly - if your output was fine before you added it, you didn't need it. The one real trap is applying it where it isn't needed and "fixing" a workflow that wasn't broken.