▢ Empty Latent Image

The vanilla Empty Latent Image node asks for raw width and height. This one asks you to pick from 23 labeled presets instead - "Portrait | Manga Cover (4:5) - 912x1144," "Mobile | Vertical Story (9:16) - 768x1360" - so you don't have to already know which numbers Illustrious was actually trained on.

Why the presets matter

Illustrious's native training resolution is 1024×1024, and later versions (v1.0/v1.1) handle up to 1536×1536 as long as width times height stays under roughly 1536² and both dimensions are multiples of 32 - that's a real constraint from the model's own training, not an arbitrary UI limit. The common working set the community settled on is 1024×1024 square, 832×1216 portrait, 1216×832 landscape, and 1344×768 widescreen. This node's preset list is built around exactly that set, plus a batch of niche ratios (character sheets, VN CG panels, manga covers) for specific use cases. enable_native_resolution (on by default) snaps your pick to those model-friendly dimensions automatically, so even if you eyeball a size, it won't hand your sampler something the model has never seen at inference.

The version-detection angle

This is the one field worth pausing on: model_version and the optional model input. Leave model_version on auto and don't plug anything into model, and the node falls back to guessing EPS vs. VPred from resolution alone - a rough heuristic. Plug your actual MODEL output in, and it can infer the real answer from the checkpoint itself. This matters because v-prediction (vpred) checkpoints - NoobAI vpred, and the closed Illustrious v3.x vpred line - need meaningfully different downstream handling than standard epsilon-prediction models: different CFG range, often a different sampler family entirely. If you're running a vpred build, plug the model in rather than trusting the guess.

noise_pattern defaults to illustrious_optimized, described as a noise flavor "tuned for Illustrious anime/illustration" as opposed to standard, high_frequency, or low_frequency - leave it unless you have a specific reason to chase a different texture.

The inputs and outputs that matter

• resolution and batch_size are required - resolution is the whole point of the node.
• model (optional) - plug it in whenever you're on a vpred checkpoint; it's the difference between this node guessing and knowing.
• optimization_mode - Quality / Speed / Compatibility, self-explanatory tradeoffs.
• seed - a base seed for noise generation, separate from whatever seed your sampler uses.

Outputs: latent feeds your sampler directly, resolution_info is a diagnostic string reporting what was actually generated - useful for confirming the preset resolved to the dimensions you expected.

Installing it

ComfyUI Manager: search "EasyIllustrious," install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/regiellis/ComfyUI-EasyIllustrious
pip install -r ComfyUI-EasyIllustrious/requirements.txt

No downloads beyond the pack itself - this node just generates an empty tensor at the right shape.

Common issues

Output looks worse at a preset you picked for its aspect ratio, not its scale. Some presets go past the recommended 1536×1536 ceiling. If quality drops off at a given resolution, drop back to a smaller preset with the same aspect ratio and use IllustriousLatentUpscale for the rest of the size increase instead of generating that large in one pass.

Vpred checkpoint producing static or nonsense. That's a classic sign the sampling pipeline downstream is running in epsilon mode against a vpred checkpoint. Check that model is actually wired into this node so version detection has something real to work from, and that your sampler is set up for vpred (v_prediction sampling, appropriate CFG).