SizorCloud/ideogram4-text-encoder-abliterated-merged

Ideogram 4 — Qwen3-VL text encoder (abliterated), single-file merged build

A single-file re-packaging of noelthekke/ideogram4-text-encoder-abliterated: the four sharded model-0000X-of-00004.safetensors are merged into one qwen3vl_text_encoder_abliterated.safetensors (398 tensors, BF16, ~15 GB), exactly as the shards load in memory. No weights, quantization or tensors are changed by the merge — it is a format repack only; the original sharded repo stays the source of truth for the weights.

Why

Local UIs that load one file per module (e.g. sd-webui-forge-classic 2.30 with native Ideogram 4.0 support) cannot pick a sharded directory in their Text-Encoder dropdown — they need one file. This repo provides exactly that file, plus the tokenizer/config needed to load it.

Files

• qwen3vl_text_encoder_abliterated.safetensors — merged text tower (BF16, flat key layout: embed_tokens.weight, model.layers.* … no visual/lm_head)
• config.json, tokenizer.*, vocab.json, merges.txt, added_tokens.json, special_tokens_map.json, chat_template.jinja — copied verbatim from the original repo
• LICENSE.md, NOTICE.txt — license and notice copied verbatim from the original repo

Usage (Forge 2.30, Ideogram 4.0)

Download qwen3vl_text_encoder_abliterated.safetensors into models/text_encoder/ and select it in the Text Encoder dropdown together with the Flux2-family VAE in the VAE dropdown, then use the Ideogram4 preset. The engine auto-downloads this file when no encoder is selected in the UI.

Credits

All weights and the tokenizer belong to the original repo author (noelthekke) and the underlying model to Ideogram Research / Qwen teams; license terms are unchanged — see LICENSE.md.