Empty HiDream-O1 Latent Image

Empty HiDream-O1 Latent Image is the blank canvas for HiDream-O1, and it breaks the rules of every empty latent you've used before. Look at the shape it builds: [batch_size, 3, height, width]. Three channels at full resolution. No 8× downscale, no 16× downscale, no deep channel stack. That's because HiDream-O1 is a pixel-space model - it denoises actual pixels directly, no VAE encoder, no VAE decoder, no latent compression at all. The "latent" here is the image itself, and this node just hands the sampler a block of zeros the right size.

The inputs: width and height (both default 2048, step 32, range 64–4096) and batch_size (default 1, max 64). The defaults aren't arbitrary - the model was trained at roughly 4 megapixels, and the node's own description is unusually blunt about it: go below that and you're off-distribution, and quality regresses noticeably. The list of trained resolutions is 2048×2048, 2304×1728, 1728×2304, 2560×1440, 1440×2560, 2496×1664, 1664×2496, 3104×1312, 1312×3104, 2304×1792 and 1792×2304. If you want HiDream's best, pick one of those rather than a random aspect ratio.

What this means in practice

Because there's no VAE, the workflow around this node looks alien to SD/Flux habits: no VAE loader, no VAE encode, no VAE decode. The sampler's output is the finished image. People routinely report HiDream workflows as "broken" for missing those pieces - they're not broken, there's just nothing to load.

The flip side is cost. Pixel-space inference pays compute proportional to pixel count, which is the whole reason VAEs exist. HiDream-O1 needs minutes for a 4MP image where a latent-space model like Z-Image does a comparable one in seconds - that's the documented trade the community hit when HiDream landed. And the blank itself is huge: a 2048×2048×3 float tensor is ~48MB before you've generated anything, and the denoising runs on that full-res tensor at every step. Budget VRAM accordingly; this is a node that wants a big card, and the description's resolution warnings are the model telling you not to cheap out on canvas size either.

One-line summary of the gotchas

Keep the dimensions to a trained resolution (or close to it), don't go hunting for a VAE that doesn't exist, and expect slowness as a feature of the architecture, not a bug in your setup. It ships with ComfyUI core; the HiDream-O1 checkpoint is the only download you need alongside it.