Experiment with Huihui Qwen3-VL text encoder

app.py

@@ -13,6 +13,7 @@ import gradio as gr
 import spaces
 import torch
 from huggingface_hub import hf_hub_download
+from transformers import AutoModel
 
 from diffusers import Ideogram4Pipeline
 from diffusers.quantizers.bitsandbytes.bnb_quantizer import BnB4BitDiffusersQuantizer
@@ -30,6 +31,7 @@ BnB4BitDiffusersQuantizer.check_quantized_param_shape = _check_quantized_param_s
 
 MODEL_ID = "ideogram-ai/ideogram-4-nf4"
 AOTI_REPO = "multimodalart/i4-block-aoti"
+TEXT_ENCODER_ID = os.environ.get("TEXT_ENCODER_ID", "huihui-ai/Huihui-Qwen3-VL-8B-Instruct-abliterated")
 MAX_SEED = 2**31 - 1
 HF_TOKEN = os.environ.get("HF_TOKEN")
 
@@ -96,7 +98,23 @@ def normalize_caption(raw_caption):
 
 
 t = time.perf_counter()
-pipe = Ideogram4Pipeline.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16, token=HF_TOKEN)
+if TEXT_ENCODER_ID:
+    text_encoder = AutoModel.from_pretrained(
+        TEXT_ENCODER_ID,
+        torch_dtype=torch.bfloat16,
+        token=HF_TOKEN,
+        low_cpu_mem_usage=True,
+    )
+    print(f"[model] using alternate text encoder: {TEXT_ENCODER_ID}", flush=True)
+else:
+    text_encoder = None
+
+pipe = Ideogram4Pipeline.from_pretrained(
+    MODEL_ID,
+    text_encoder=text_encoder,
+    torch_dtype=torch.bfloat16,
+    token=HF_TOKEN,
+)
 pipe.transformer.dequantize()
 pipe.unconditional_transformer.dequantize()
 pipe.to("cuda")
@@ -180,7 +198,8 @@ textarea { font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, monospace
 with gr.Blocks(theme=gr.themes.Citrus(), title="Ideogram 4 JSON Lab", css=CSS) as demo:
     gr.Markdown(
         "# Ideogram 4 JSON Lab\n"
-        "Direct structured JSON caption input for Ideogram 4. No remote magic prompt, no local Qwen prompt upsampling."
+        "Direct structured JSON caption input for Ideogram 4. No remote magic prompt, no local Qwen prompt upsampling.\n\n"
+        f"Text encoder: `{TEXT_ENCODER_ID or 'ideogram-ai/ideogram-4-nf4 bundled Qwen3-VL'}`"
     )
 
     with gr.Row():