<h2 id="layerwise-llm-for-ideogram4">Layerwise LLM for ideogram4</h2>This version will only work with the ideogram official pipeline not in comfyNOTE: Ideogram4 Naming is the same as QWEN with the prefix "model." removed (Same Block Structure) COMFY in there infinite wisdom reordered the blocks - FUCK COMFY, I really don't feel like trying to reverse engineer the custom shit they did again. This is a trained version of the Qwen3-VL-8B-Instruct model.<ul><li>Full training of entire model (TF32/BF16 Full Adam)</li><li>Layerwise training in FP32 of target layers used by ideogram</li></ul>

Files with SHA256: ac82ca227805f8c35d5dd77968cc4b51008d875564d2b6b9a731c0d53f2c764e