Model Scale HiDream

HiDream-I1 was April 2025's big MIT-licensed challenger to Flux: excellent prompt adherence, far too heavy for most hardware, and the node ecosystem around it has always been thin. If you're running the Full, Dev, or Fast variants - 17B of latent DiT, so you're already on serious hardware or a good quant - Model Scale HiDream is the per-block scaler for it. Same idea as the SDXL one in this pack, aimed at HiDream's particular split: 13 dense double-stream blocks up front and 32 single-stream blocks after.

The mechanism is the pack's standard scale_model_by_prefix: clone, match weights to the longest widget prefix, scale by weight × scale, leave everything at 1.0 by default. 0.0 zeroes a block out, values above 1.0 amplify. model in, one MODEL out.

The inputs that matter

• x_embedder, t_embedder, caption_projection - the input path: image tokens, timestep, and caption. caption_projection is the one worth knowing, because HiDream's famous prompt adherence flows through it. Scaling it up amplifies how hard the model follows the caption; too high and the image starts to warp around the text.
• double_stream_blocks.0–double_stream_blocks.12 - the early joint-attention blocks. This is where prompt semantics are fused into the image stream. If you want more prompt influence, this is usually the region.
• single_stream_blocks.0–single_stream_blocks.31 - the long tail, which refines detail toward the output. These are the style/detail blocks; scaling them down softens and loosens the render.

That double/single split is the whole reason a HiDream-specific node exists - a generic block scaler wouldn't know HiDream has 13 and 32 of each, or that caption_projection is a distinct, meaningful layer.

How to use it

Pick one region per test run. The classic starting experiment: nudge caption_projection to ~1.2 and the first few double-stream blocks to ~1.1, sample a few seeds, and compare prompt adherence against the unmodified model. Or, for the "too heavy, make it lazier" direction, drop the last third of single-stream blocks to 0.7–0.9 to shave some generation cost and see what quality you lose. Keep changes small - a 17B model already tuned hard for prompt following doesn't tolerate big scale swings gracefully.

Install

Search Easygoing in the ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/easygoing0114/ComfyUI-easygoing-nodes.git

Restart ComfyUI. No pip extras; requires a current ComfyUI with the V3 node API.

What to watch

HiDream-I1's weight is its problem - holding the model plus its clone in memory is the practical ceiling, and on 16GB cards you'll want the quantized variant regardless. Also, this node assumes the standard HiDream block counts (double 0–12, single 0–31); if your checkpoint was re-cut or fused, verify with Key Name Inspector before trusting a slider. And as always with these scale nodes: the effect lives in the graph, so save the workflow - or route through the pack's save-with-original pair if you want the scaled model as a reusable file.