# Fix: Ideogram 4.

> Source: <https://promptcube3.com/en/threads/5004/>
> Published: 2026-08-04 22:27:24+00:00

# Fix: Ideogram 4.

## The Actual Error

When invoking the Ideogram node with spatial prompts like `"A sign reading 'OPEN' hanging above a door"`

, the server log spat this out:

```
[ERROR] Text encoding failed: expected hidden_states dim [1, 256, 1024], got [1, 203, 1024]
[WARN]  Bounding box projection skipped — token alignment broken
```

The node loads the model fine, but the text encoder's token padding length doesn't match what the unet expects. This kills spatial bounding boxes and any prompt with more than ~4 words.

## Diagnosis

I compared the custom node's `model_config.json`

against Ideogram's reference implementation. Two issues:

1. The node forces `max_position_embeddings: 256`

while the local checkpoint was trained with `203`

2. The tokenizer pads to the config value instead of detecting the actual sequence length

Swapping to dynamic padding based on the real token count fixed the dimension mismatch. I also had to patch the bounding box sampler to re-align tokens after truncation.

## Steps That Worked

1. Clone the node into `ComfyUI/custom_nodes/`

2. Edit `nodes/ideogram_node.py`

— replace the static padding with:

```
# Before (broken)
encoded = tokenizer(prompt, padding="max_length", max_length=256, truncation=True)

# After (fixed)
encoded = tokenizer(prompt, padding="longest", truncation=True, return_tensors="pt")
max_len = encoded.input_ids.shape[1]
# Resize position embeddings if needed
model.resize_position_embeddings(max_len)
```

3. In `model_config.json`

, set `"max_position_embeddings": 203`

to match the checkpoint

4. Restart ComfyUI — the node now preserves character counts and bounding box alignment

## What Still Isn't Perfect

The local run is noticeably slower than the cloud API (batch size 1 on a 4090 takes ~18s vs ~6s). Also, I haven't cracked exact spatial prompt syntax yet — `"text above image"`

sometimes renders below. If anyone has the correct token-bounding format for local Ideogram 4.0, I'm all ears.

For now, this patch delivers clean text rendering and functional bounding boxes on consumer hardware. I'll push the fork once I clean up the sampler fix.

[Next XXO-Bench →](/en/threads/5001/)

[a library of Claude prompt techniques](http://154.12.95.112/), with plenty of directly applicable cases.