Ideogram4 LoRA Block Weight

Ideogram 4 is a 34-layer single-stream DiT - text and image tokens flow through one unified stack, and it's the model that made people finally stop talking about CLIP. But its LoRAs work like every other LoRA: one global strength, applied everywhere, whether the subject needed it in the attention layers or just the final projection. Ideogram4 LoRA Block Weight is the node that gives you per-block and per-layer-type control instead.

It's the configurator half of the ComfyUI-LoRA-Block-Filter pack (by Danrisi, the Anima "UltraReal" fine-tuner). It doesn't touch any models - it builds a weight configuration and hands it to the sibling Ideogram4 LoRA Filter (Blocks) node, which does the loading and patching. A slightly unusual pick for this pack since Ideogram 4's ecosystem was stalled out by its license within a month of release, but if you're running the fp8 or nf4 checkpoints locally and training your own LoRAs, this is the tool for controlling what they do.

Inputs

• preset - ten options. Beyond the shared all_enabled, character_detail, no_style_leak, face_focused, style_transfer, attn_only, mlp_only and custom, this node adds two it doesn't share with the Anima/Flux versions: early_blocks (which drops final_layer only) and late_blocks (which drops the input and LLM-conditioning projections). The no_style_leak preset here zeroes llm_cond_proj and pulls adaln down - on Ideogram 4, prompt/style conditioning comes through the LLM text path, so that's where leak suppression lives.
• The seven layer-type sliders, each with the author's own tooltip to lean on:
  • attn_qkv - attention QKV projection
  • attn_out - attention output projection
  • mlp - the SwiGLU MLP (w1/w2/w3)
  • adaln - AdaLN modulation (timestep conditioning)
  • input_proj - input projection (latent→hidden)
  • llm_cond_proj - LLM conditioning projection (text→hidden)
  • final_layer - final output layer
• block_vector - 34 comma-separated multipliers, one per layer, applied on top of the layer-type weights (the tooltip says exactly that).

Two levels of control, same as the Flux version of this node: coarse sliders for layer type, then a per-layer vector for surgical edits. Want attention on early layers but not late ones? You can do that with the vector alone.

Output

A single block_weights output of type IDEOGRAM4_BLOCK_WEIGHTS, which plugs into the block_weights input on Ideogram4 LoRA Filter (Blocks). Nothing else passes through this node.

Install

ComfyUI Manager handles it under the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/DanrisiUA/ComfyUI-LoRA-Block-Filter

Restart, and you'll find it under Smart LoRA. No requirements.txt, no model downloads - this pack rides entirely on ComfyUI's built-in LoRA code.

One thing to know

Ideogram 4 shipped no bf16 checkpoint, only fp8 and nf4, and LoRA support in ComfyUI for it is recent - you want an up-to-date ComfyUI for any of this to make sense. And because the filter only recognizes ai-toolkit-style key names (diffusion_model.layers.N.*), a LoRA trained with unusual naming won't match the block logic and every key will pass through at full strength. Check the filter's info output to confirm it's actually filtering.