Model Scale ERNIE Image

ERNIE-Image is Baidu's 8B Apache-2.0 model - the one that genuinely impressed people with structured layout and text-in-image before the community's attention moved on. If you're one of the people still using it, Model Scale ERNIE Image is the fine-grained scaler for it: one slider per layer and per attention component, covering all 36 transformer blocks of ernie-image.safetensors. This is the most granular scaler in the pack, and it exists because ERNIE-Image's strength - following structured prompts - isn't spread evenly across its weights.

The mechanism is the same as every scale node here: clone the model, match each weight to the longest matching widget prefix, and scale by weight × scale via add_patches. 1.0 leaves a layer untouched, 0.0 zeroes it out, values above 1.0 amplify. All sliders default to 1.0, so the node is inert until you move something. Input is a model, output is a single scaled MODEL for your sampler.

The inputs, grouped

The schema is huge - hundreds of floats - because every layer is expanded into its sub-components. But they organize cleanly:

• layers.N - the per-block master slider. Start here; the sub-sliders are for when you need finer control.
• layers.N.self_attention.to_q / .to_k / .to_v / .to_out - the individual attention projection matrices inside each block. This is the rare level of control: you can weaken the query projections across all blocks without touching the MLPs, which is a very different surgical move than scaling whole blocks.
• layers.N.mlp.gate_proj / .up_proj / .linear_fc2 - the feed-forward path, the part most responsible for factual/layout recall in these architectures.
• layers.N.adaLN_mlp_ln / .adaLN_sa_ln - the adaptive normalization layers that modulate each block.
• Top-level: x_embedder, text_proj, time_embedding, adaLN_modulation, final_norm, final_linear.

What to actually try

If you're chasing ERNIE-Image's text-rendering quality, the community's finding is that its competence is real but front-loaded - the mid-to-late blocks and the MLP path carry a lot of it. A reasonable first experiment: amplify mlp sub-components in the middle third (layers ~12–24) slightly above 1.0 and see if layout fidelity sharpens, then dial back to taste. Zeroing whole late blocks usually just degrades output. And if the grid-pattern artifact that people noted with ERNIE-Image is bothering you, experimenting with the attention projections in the early layers is a more targeted lever than a global scale.

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; needs a current ComfyUI with the V3 node API for the pack to register.

Honest warning

This is a scaler for people already invested in ERNIE-Image, and the sub-component granularity is genuinely useful there - but the node only makes sense if the widget prefixes match your model's actual keys. Run it once with the pack's Key Name Inspector wired in parallel if you're not sure; a mismatch shows up as sliders that do nothing, and that's a wasted afternoon otherwise. Scaling is also in-graph only - save the workflow, or route through the pack's save-with-original nodes if you want a file.