fix hf telemetry and add prxpixel

Signed-off-by: Vladimir Mandic <[email protected]>

.github/skills/port-model/SKILL.md

@@ -179,9 +179,9 @@ Add only what the model actually needs.
 
 Reference catalog touchpoints are also required for model ports intended to appear in SD.Next model references.
 
-- `data/reference.json` for `base`
+- `data/reference-base.json` for `base`
 - `data/reference-cloud.json` for `cloud`
-- `data/reference-quant.json` for `quant`
+- `data/reference-quantized.json` for `quant`
 - `data/reference-distilled.json` for `distilled`
 - `data/reference-nunchaku.json` for `nunchaku`
 - `data/reference-community.json` for `community`

.github/skills/reference-catalog/SKILL.md

@@ -21,9 +21,9 @@ Use this skill to audit and update SD.Next model reference catalogs using a phas
 
 ## Catalog Files In Scope
 
-- `data/reference.json` (base)
+- `data/reference-base.json` (base)
 - `data/reference-cloud.json`
-- `data/reference-quant.json`
+- `data/reference-quantized.json`
 - `data/reference-distilled.json`
 - `data/reference-nunchaku.json`
 - `data/reference-community.json`

CHANGELOG.md

@@ -16,6 +16,8 @@ And **SDNQ** improvements: now with *NPU* support and its own native *attention*
   - [Krea 2](https://www.krea.ai/blog/krea-2-image-model) in *base* and *turbo* (distilled) variants  
     both *base* and *turbo* are available in both full *bf16* and *sdnq* pre-quantized variants  
     K2 is a 12.9B single-stream flow-matching DiT and using a Qwen3-VL-4B text encoder  
+  - [Photoroom PRXPixel](https://huggingface.co/Photoroom/prxpixel-t2i) pixel-space PRX variant using a Qwen3-VL text encoder and flow-matching scheduler  
+    supports *direct RGB* generation without a VAE and uses a *1024px* default sample size  
   - **Microsoft Lens** got unpublished, but we still got a mirror
 - **Features**
   - **SDNQ-Attention**  

TODO.md

@@ -18,6 +18,8 @@
 
 ### Unassigned
 
+- Implement [nvvfx](https://pypi.org/project/nvidia-vfx/)
+- Implement [pruna](https://github.com/PrunaAI/pruna)
 - Processing -> Video capabilities
   - `RIFE` in processing
   - `SeedVR2` in processing
@@ -36,7 +38,6 @@
 
 ### OnHold
 
-- Implement [pruna](https://github.com/PrunaAI/pruna), @vladmandic, pending support for transformers 5.5
 - LoRA add OMI format support for SD35/FLUX.1, on-hold
 - Remote Text-Encoder support, sidelined for the moment
 - Multi-user support

data/reference-base.json

@@ -794,6 +794,12 @@
     "preview": "Photoroom--prx-1024-t2i-beta.jpg",
     "size": 20.7
   },
+  "Photoroom PRXPixel": {
+    "path": "Photoroom/prxpixel-t2i",
+    "desc": "PRXPixel is a 7-billion-parameter pixel-space text-to-image model that denoises RGB images directly using a Qwen3-VL text encoder and flow-matching scheduler.",
+    "preview": "Photoroom--prxpixel-t2i.jpg",
+    "size": 25.0
+  },
   "ZAI GLM-Image": {
     "path": "zai-org/GLM-Image",
     "preview": "zai-org--GLM-Image.jpg",

modules/modeldata.py

@@ -111,6 +111,8 @@ def get_model_type(pipe):
         model_type = 'nextstep'
     elif 'XOmni' in name or 'X-Omni' in name:
         model_type = 'x-omni'
+    elif 'PRXPixel' in name or 'PRXPixelPipeline' in name:
+        model_type = 'prxpixel'
     elif 'Photoroom' in name:
         model_type = 'prx'
     elif 'LongCat' in name:

modules/sd_detect.py

@@ -141,6 +141,8 @@ def guess_by_name(fn, current_guess):
         new_guess = 'Stable Diffusion XL'
     elif 'stable-video-diffusion' in fn.lower():
         new_guess = 'StableVideoDiffusion'
+    elif 'prx-pixel' in fn.lower() or 'prxpixel' in fn.lower():
+        new_guess = 'PRXPixel'
     elif 'prx-' in fn.lower():
         new_guess = 'PRX'
     elif 'gemini-' in fn.lower() and 'image' in fn.lower():

modules/sd_hijack_hfhub.py

@@ -8,6 +8,19 @@ orig_http_get = None
 orig_xet_get = None
 
 
+def clean_user_agent(headers):
+    ua = headers.get('user-agent', None)
+    if ua is None:
+        return headers
+    ua = ua.replace('unknown/None;', 'sdnext/0.0.0;')
+    if 'quant_config' in ua:
+        ua = ua.split('quant_config')[0]
+    ua = ua.rstrip(' ;')
+    # ua += '; telemetry/off'
+    headers['user-agent'] = ua
+    return headers
+
+
 def http_get_hijack(*args, **kwargs):
     from modules.shared import state, opts
     if len(args) > 0 and isinstance(args[0], str) and args[0].endswith(".json"):
@@ -15,6 +28,8 @@ def http_get_hijack(*args, **kwargs):
     jobid = state.begin('Download')
     fn = kwargs.get("displayed_filename", None)
     size = kwargs.get("expected_size", None)
+    if 'headers' in kwargs:
+        kwargs['headers'] = clean_user_agent(kwargs['headers'])
     if fn and not fn.endswith(".json") and size is not None and size > 10240:
         log.debug(f'Download: type=http mode="{opts.hf_transfer_mode}" fn="{fn}" size={size}')
     debug(f'Download start: type=http args={args} kwargs={kwargs}')
@@ -33,6 +48,8 @@ def xet_get_hijack(*args, **kwargs):
     jobid = state.begin('Download')
     fn = kwargs.get("displayed_filename", None)
     size = kwargs.get("expected_size", None)
+    if 'headers' in kwargs:
+        kwargs['headers'] = clean_user_agent(kwargs['headers'])
     if fn and not fn.endswith(".json"):
         log.debug(f'Download: type=xet mode="{opts.hf_transfer_mode}" fn="{fn}" size={size}')
     debug(f'Download start: type=xet args={args} kwargs={kwargs}')

modules/sd_models.py

@@ -524,7 +524,11 @@ def load_diffuser_force(detected_model_type: str, checkpoint_info: CheckpointInf
             from pipelines.model_google import load_nanobanana
             sd_model = load_nanobanana(checkpoint_info, diffusers_load_config)
             allow_post_quant = False
-        elif model_type in ['PRX']:
+        elif model_type == 'PRXPixel':
+            from pipelines.model_prx_pixel import load_prx_pixel
+            sd_model = load_prx_pixel(checkpoint_info, diffusers_load_config)
+            allow_post_quant = False
+        elif model_type == 'PRX':
             from pipelines.model_prx import load_prx
             sd_model = load_prx(checkpoint_info, diffusers_load_config)
             allow_post_quant = False

modules/shared_items.py

@@ -54,6 +54,7 @@ pipelines = {
     'OmniGen': getattr(diffusers, 'OmniGenPipeline', None),
     'PixArtAlpha': getattr(diffusers, 'PixArtAlphaPipeline', None),
     'PixArtSigma': getattr(diffusers, 'PixArtSigmaPipeline', None),
+    'PRXPixel': getattr(diffusers, 'PRXPixelPipeline', None),
     'Qwen': getattr(diffusers, 'QwenImagePipeline', None),
     'Sana': getattr(diffusers, 'SanaPipeline', None),
     'WanAI': getattr(diffusers, 'WanPipeline', None),

pipelines/model_prx_pixel.py

@@ -0,0 +1,39 @@
+import diffusers
+import transformers
+from modules import shared, devices, sd_models, model_quant, sd_hijack_te
+from modules.logger import log
+from pipelines import generic
+
+
+def load_prx_pixel(checkpoint_info, diffusers_load_config=None):
+    if diffusers_load_config is None:
+        diffusers_load_config = {}
+    repo_id = sd_models.path_to_repo(checkpoint_info)
+    sd_models.hf_auth_check(checkpoint_info)
+
+    load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config, allow_quant=False)
+    log.debug(
+        f'Load model: type=PRXPixel repo="{repo_id}" config={diffusers_load_config} '
+        f'offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}'
+    )
+
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.PRXTransformer2DModel, load_config=diffusers_load_config)
+    text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3VLTextModel, load_config=diffusers_load_config)
+
+    if repo_id is None or repo_id.lower() == 'none':
+        return None
+
+    pipe = diffusers.PRXPixelPipeline.from_pretrained(
+        repo_id,
+        transformer=transformer,
+        text_encoder=text_encoder,
+        cache_dir=shared.opts.diffusers_dir,
+        **load_args,
+    )
+
+    del text_encoder
+    del transformer
+    sd_hijack_te.init_hijack(pipe)
+
+    devices.torch_gc()
+    return pipe