DiTCondLabelSelect

This is how you "prompt" the original Facebook Research DiT model - and it's genuinely not a text box. DiT was trained purely on ImageNet classification, so instead of typing a description, you pick one of the actual 1000 ImageNet class names from a dropdown: tench, Tinca tinca, goldfish, Carassius auratus, great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias, hammerhead, hammerhead shark, all the way through categories most people have never had reason to think about as image-generation prompts. If you're used to natural-language prompting, this node is a small, genuinely interesting reminder of what image generation looked like before text conditioning became the default - a single fixed vocabulary, exhaustively enumerated, no ambiguity about what any given label means.

Inputs. Two, both required. model is your loaded DiT model from DitCheckpointLoader. label_name is the class dropdown - all 1000 ImageNet categories, verbatim class names including the taxonomic-sounding secondary terms (Tinca tinca alongside tench, for instance) that come straight from the original dataset labels.

Output. class, typed as CONDITIONING. Wire it into your KSampler's positive input, with DiTCondLabelEmpty's output going to negative - the pack's own README is explicit that negative specifically needs the empty-label node, not a second real class.

What you can and can't combine. The README notes ConditioningCombine nodes "should" work for blending multiple class labels together - reasonable if you want something between two categories. What doesn't work: the "area" conditioning nodes that let you paint different prompts onto different regions of the image elsewhere in ComfyUI. DiT can't handle the dynamic input dimensions those require, so don't build a workflow expecting spatial control over which class shows up where.

Installing the pack. ComfyUI Manager: search "Extra Models for ComfyUI." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI_ExtraModels

then pip install -r requirements.txt and restart.

Getting a full DiT workflow running. You need weights (original from facebookresearch/DiT or the author's converted FP16 safetensors) loaded via DitCheckpointLoader, this node for positive conditioning, DiTCondLabelEmpty for negative, a standard SD1.5 VAE and EmptyLatentImage (DiT shares SD1.5's latent space, no special VAE needed), and output capped at 256 or 512px depending on which image_size you picked at load time.

Realistic expectations. This isn't going to compete with a modern text-to-image model on flexibility - you get one label per image (or a blend, via ConditioningCombine), fixed to whatever taxonomy ImageNet happened to use in 2009. Where it's genuinely worth having is as a fast, well-understood baseline: if you're studying how DiT architectures behave, or want a clean reference point before comparing against PixArt or Sana (both descended from this same architecture with text conditioning layered on top), this is the least confounded way to see the base architecture at work.