add hidream

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -1,13 +1,30 @@
 # Change Log for SD.Next
 
-## Update for 2025-04-08
+## TODO for 2025-04-10
 
+- HiDream requires sidebranch:  
+  > `pip install git+https://github.com/hlky/diffusers@hidream`  
+  > `./webui.sh --debug --experimental`  
+- FlashAttn wiki  
+
+## Update for 2025-04-10
+
+- **Models**  
+  - [HiDream-I1](https://huggingface.co/HiDream-ai/HiDream-I1-Full) in fast, dev and full variants!  
+    new absolutely massive image generative foundation model with **17B** parameters  
+    and 4 (!?) text-encoders: *clip-l, clip-g, t5-1.1-xxl, llama-3.1-8b-instruct* for total of **8.3B** parameters  
+    simply select from *networks -> models -> reference*  
+    due to size (over 25B params in 58GB), offloading and on-the-fly quantization are pretty much a necessity  
+    difference between variants is recommended number of steps: *fast=16, dev=28, full=50*  
+    hidream is compatible with flowmatching samplers and with taesd live-preview  
+    *note* HiDream-I1 requires `flash-attn` to be installed  
+    > REQUIRES SIDE-BRANCH
 - **Features**  
   - Custom model loader  
-    can be used to load any known diffusion model with default or custom model components  
-    in models -> custom tab  
-    see docs for details: <https://vladmandic.github.io/sdnext-docs/Loader/>  
-  - Pipe: [SoftFill](https://github.com/zacheryvaughn/softfill-pipelines)  
+      can be used to load any known diffusion model with default or custom model components  
+      in models -> custom tab  
+      see docs for details: <https://vladmandic.github.io/sdnext-docs/Loader/>  
+    - Pipe: [SoftFill](https://github.com/zacheryvaughn/softfill-pipelines)  
 - **Caching**  
   - add `TeaCache` support to *Flux, CogVideoX, Mochi, LTX*  
   - add `FasterCache` support to *WanAI, LTX* (other video models already supported)  
@@ -23,9 +40,12 @@
   - ZLUDA: add more GPUs to recognized list  
     select in scripts, available for sdxl in inpaint model  
   - LoRA: add option to force-reload LoRA on every generate  
+  - settings: add **Model options** sections as placeholder for per-model settings
+  - video: update *LTXVideo-0.9.5* pipeline  
+  - te loader: allow free-form input in which case sdnext will attempt to load it as hf repo  
   - diag: add get-server-status to UI generate context menu  
   - diag: memory monitor detect gpu swapping  
-  - vide: update LTXVideo-0.9.5 pipeline  
+  - use [hf-xet](https://huggingface.co/blog/xet-on-the-hub) for huggingface downloads where possible  
 - **Changes**  
   - params: Reset default guidance-rescale from 0.7 to 0.0  
   - progress: add additional fields to progress API  

html/reference.json

@@ -333,7 +333,29 @@
     "preview": "Alpha-VLLM--Lumina-Image-2.0.jpg",
     "skip": true,
     "extras": "sampler: Default"
-  }, 
+  },
+
+  "HiDream-I1 Fast": {
+    "path": "HiDream-ai/HiDream-I1-Fast",
+    "desc": "HiDream-I1 is a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds.",
+    "preview": "HiDream-ai--HiDream-I1-Fast.jpg",
+    "skip": true,
+    "extras": "sampler: Default"
+  },
+  "HiDream-I1 Dev": {
+    "path": "HiDream-ai/HiDream-I1-Dev",
+    "desc": "HiDream-I1 is a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds.",
+    "preview": "HiDream-ai--HiDream-I1-Fast.jpg",
+    "skip": true,
+    "extras": "sampler: Default"
+  },
+  "HiDream-I1 Full": {
+    "path": "HiDream-ai/HiDream-I1-Full",
+    "desc": "HiDream-I1 is a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds.",
+    "preview": "HiDream-ai--HiDream-I1-Fast.jpg",
+    "skip": true,
+    "extras": "sampler: Default"
+  },
 
   "Kwai Kolors": {
     "path": "Kwai-Kolors/Kolors-diffusers",

modules/model_cogview.py

@@ -29,16 +29,16 @@ def load_cogview3(checkpoint_info, diffusers_load_config={}):
     repo_id = sd_models.path_to_repo(checkpoint_info.name)
     shared.log.debug(f'Load model: type=CogView3 model="{checkpoint_info.name}" repo="{repo_id}" offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype}')
 
-    diffusers_load_config, quant_args = load_common(diffusers_load_config, module='Model')
+    load_args, quant_args = load_common(diffusers_load_config, module='Transformer')
     transformer = diffusers.CogView3PlusTransformer2DModel.from_pretrained(
         repo_id,
         subfolder="transformer",
         cache_dir=shared.opts.diffusers_dir,
-        **diffusers_load_config,
+        **load_args,
         **quant_args,
     )
 
-    diffusers_load_config, quant_args = load_common(diffusers_load_config, module='TE')
+    load_args, quant_args = load_common(diffusers_load_config, module='TE')
     text_encoder = transformers.T5EncoderModel.from_pretrained(
         repo_id,
         subfolder="text_encoder",
@@ -47,12 +47,13 @@ def load_cogview3(checkpoint_info, diffusers_load_config={}):
         **quant_args,
     )
 
+    load_args, quant_args = load_common(diffusers_load_config, module='Transformer')
     pipe = diffusers.CogView3PlusPipeline.from_pretrained(
         repo_id,
         text_encoder=text_encoder,
         transformer=transformer,
         cache_dir=shared.opts.diffusers_dir,
-        **diffusers_load_config,
+        **load_args,
     )
     devices.torch_gc()
     return pipe
@@ -62,7 +63,7 @@ def load_cogview4(checkpoint_info, diffusers_load_config={}):
     repo_id = sd_models.path_to_repo(checkpoint_info.name)
     shared.log.debug(f'Load model: type=CogView4 model="{checkpoint_info.name}" repo="{repo_id}" offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype}')
 
-    diffusers_load_config, quant_args = load_common(diffusers_load_config, module='Model')
+    load_args, quant_args = load_common(diffusers_load_config, module='Transformer')
     transformer = diffusers.CogView4Transformer2DModel.from_pretrained(
         repo_id,
         subfolder="transformer",
@@ -71,21 +72,22 @@ def load_cogview4(checkpoint_info, diffusers_load_config={}):
         **quant_args,
     )
 
-    diffusers_load_config, quant_args = load_common(diffusers_load_config, module='TE')
+    load_args, quant_args = load_common(diffusers_load_config, module='TE')
     text_encoder = transformers.AutoModelForCausalLM.from_pretrained(
         repo_id,
         subfolder="text_encoder",
         cache_dir=shared.opts.diffusers_dir,
-        **diffusers_load_config,
+        **load_args,
         **quant_args,
     )
 
+    load_args, quant_args = load_common(diffusers_load_config, module='Model')
     pipe = diffusers.CogView4Pipeline.from_pretrained(
         repo_id,
         text_encoder=text_encoder,
         transformer=transformer,
         cache_dir=shared.opts.diffusers_dir,
-        **diffusers_load_config,
+        **load_args,
     )
     if shared.opts.diffusers_eval:
         pipe.text_encoder.eval()

modules/model_hidream.py

@@ -0,0 +1,92 @@
+import time
+import transformers
+import diffusers
+from modules import shared, devices, sd_models, timer
+
+
+llama_repo = "meta-llama/Meta-Llama-3.1-8B-Instruct"
+
+
+def hijack_encode_prompt(*args, **kwargs):
+    t0 = time.time()
+    res = shared.sd_model.orig_encode_prompt(*args, **kwargs)
+    t1 = time.time()
+    timer.process.add('te', t1-t0)
+    # shared.log.debug(f'Hijack: te={shared.sd_model.text_encoder.__class__.__name__} time={t1-t0:.2f}')
+    shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model)
+    return res
+
+
+def get_args(diffusers_load_config={}, module=None):
+    from modules import model_quant, modelloader
+    modelloader.hf_login()
+    if 'torch_dtype' not in diffusers_load_config:
+        diffusers_load_config['torch_dtype'] = devices.dtype
+    if 'low_cpu_mem_usage' in diffusers_load_config:
+        del diffusers_load_config['low_cpu_mem_usage']
+    if 'load_connected_pipeline' in diffusers_load_config:
+        del diffusers_load_config['load_connected_pipeline']
+    if 'safety_checker' in diffusers_load_config:
+        del diffusers_load_config['safety_checker']
+    if 'requires_safety_checker' in diffusers_load_config:
+        del diffusers_load_config['requires_safety_checker']
+    quant_args = model_quant.create_config(module=module)
+    quant_type = model_quant.get_quant_type(quant_args)
+    if quant_type:
+        shared.log.debug(f'Load model: type=HiDream quantization module="{module}" {quant_type}')
+    return diffusers_load_config, quant_args
+
+
+def load_hidream(checkpoint_info, diffusers_load_config={}):
+    repo_id = sd_models.path_to_repo(checkpoint_info.name)
+    shared.log.debug(f'Load model: type=HiDream model="{checkpoint_info.name}" repo="{repo_id}" offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype}')
+
+    load_args, quant_args = get_args(diffusers_load_config, module='Transformer')
+    transformer = diffusers.HiDreamImageTransformer2DModel.from_pretrained(
+        repo_id,
+        subfolder="transformer",
+        cache_dir=shared.opts.hfcache_dir,
+        **load_args,
+        **quant_args,
+    )
+
+    load_args, quant_args = get_args(diffusers_load_config, module='TE')
+    text_encoder_3 = transformers.T5EncoderModel.from_pretrained(
+        repo_id,
+        subfolder="text_encoder_3",
+        cache_dir=shared.opts.hfcache_dir,
+        **load_args,
+        **quant_args,
+    )
+
+    load_args, quant_args = get_args(diffusers_load_config, module='LLM')
+    tokenizer_4 = transformers.PreTrainedTokenizerFast.from_pretrained(
+        llama_repo,
+        cache_dir=shared.opts.hfcache_dir,
+        **load_args,
+    )
+    text_encoder_4 = transformers.LlamaForCausalLM.from_pretrained(
+        llama_repo,
+        output_hidden_states=True,
+        output_attentions=True,
+        cache_dir=shared.opts.hfcache_dir,
+        **load_args,
+        **quant_args,
+    )
+
+    load_args, quant_args = get_args(diffusers_load_config, module='Model')
+    pipe = diffusers.HiDreamImagePipeline.from_pretrained(
+        repo_id,
+        text_encoder_3=text_encoder_3,
+        text_encoder_4=text_encoder_4,
+        tokenizer_4=tokenizer_4,
+        transformer=transformer,
+        cache_dir=shared.opts.diffusers_dir,
+        **load_args,
+    )
+
+    pipe.orig_encode_prompt = pipe.encode_prompt
+    pipe.encode_prompt = hijack_encode_prompt
+
+    devices.torch_gc()
+    return pipe

modules/model_sd3.py

@@ -1,7 +1,7 @@
 import os
 import diffusers
 import transformers
-from modules import shared, devices, sd_models, sd_unet, model_quant, model_tools
+from modules import shared, devices, errors, sd_models, sd_unet, model_quant, model_tools
 
 
 def load_overrides(kwargs, cache_dir):
@@ -20,6 +20,7 @@ def load_overrides(kwargs, cache_dir):
                 shared.log.debug(f'Load model: type=SD3 unet="{shared.opts.sd_unet}" fmt=gguf')
         except Exception as e:
             shared.log.error(f"Load model: type=SD3 failed to load UNet: {e}")
+            errors.display(e, 'UNet')
             shared.opts.sd_unet = 'Default'
             sd_unet.failed_unet.append(shared.opts.sd_unet)
     if shared.opts.sd_text_encoder != 'Default':
@@ -36,6 +37,7 @@ def load_overrides(kwargs, cache_dir):
                 shared.log.debug(f'Load model: type=SD3 variant="t5" te="{shared.opts.sd_text_encoder}"')
         except Exception as e:
             shared.log.error(f"Load model: type=SD3 failed to load T5: {e}")
+            errors.display(e, 'TE')
             shared.opts.sd_text_encoder = 'Default'
     if shared.opts.sd_vae != 'Default' and shared.opts.sd_vae != 'Automatic':
         try:
@@ -47,6 +49,7 @@ def load_overrides(kwargs, cache_dir):
                 shared.log.debug(f'Load model: type=SD3 vae="{shared.opts.sd_vae}"')
         except Exception as e:
             shared.log.error(f"Load model: type=SD3 failed to load VAE: {e}")
+            errors.display(e, 'VAE')
             shared.opts.sd_vae = 'Default'
     return kwargs
 
@@ -86,33 +89,6 @@ def load_missing(kwargs, fn, cache_dir):
     return kwargs
 
 
-"""
-def load_gguf(kwargs, fn):
-    ggml.install_gguf()
-    from accelerate import init_empty_weights
-    from diffusers.loaders.single_file_utils import convert_sd3_transformer_checkpoint_to_diffusers
-    from modules import ggml, sd_hijack_accelerate
-    with init_empty_weights():
-        config = diffusers.SD3Transformer2DModel.load_config(os.path.join('configs', 'flux'), subfolder="transformer")
-        transformer = diffusers.SD3Transformer2DModel.from_config(config).to(devices.dtype)
-        expected_state_dict_keys = list(transformer.state_dict().keys())
-    state_dict, stats = ggml.load_gguf_state_dict(fn, devices.dtype)
-    state_dict = convert_sd3_transformer_checkpoint_to_diffusers(state_dict)
-    applied, skipped = 0, 0
-    for param_name, param in state_dict.items():
-        if param_name not in expected_state_dict_keys:
-            skipped += 1
-            continue
-        applied += 1
-        sd_hijack_accelerate.hijack_set_module_tensor_simple(transformer, tensor_name=param_name, value=param, device=0)
-        transformer.gguf = 'gguf'
-        state_dict[param_name] = None
-    shared.log.debug(f'Load model: type=Unet/Transformer applied={applied} skipped={skipped} stats={stats} compute={devices.dtype}')
-    kwargs['transformer'] = transformer
-    return kwargs
-"""
-
-
 def load_sd3(checkpoint_info, cache_dir=None, config=None):
     repo_id = sd_models.path_to_repo(checkpoint_info.name)
     fn = checkpoint_info.path
@@ -149,6 +125,10 @@ def load_sd3(checkpoint_info, cache_dir=None, config=None):
     shared.log.debug(f'Load model: type=SD3 kwargs={list(kwargs)} repo="{repo_id}"')
 
     kwargs = model_quant.create_config(kwargs)
+    if shared.opts.model_sd3_disable_te5:
+        shared.log.debug('Load model: type=SD3 option="disable-te5"')
+        kwargs['text_encoder_3'] = None
+
     pipe = loader(
         repo_id,
         torch_dtype=devices.dtype,

modules/model_te.py

@@ -78,6 +78,10 @@ def load_t5(name=None, cache_dir=None):
                 dtype=torch.float32 if devices.dtype != torch.bfloat16 else torch.bfloat16
             )
         t5 = nncf_compress_model(t5)
+    elif '/' in name:
+        shared.log.debug(f'Load model: type=T5 repo={name}')
+        quant_config = model_quant.create_config(module='TE')
+        t5 = transformers.T5EncoderModel.from_pretrained(name, cache_dir=cache_dir, torch_dtype=devices.dtype, **quant_config)
     else:
         t5 = None
     if t5 is not None:

modules/modeldata.py

@@ -41,6 +41,8 @@ def get_model_type(pipe):
         model_type = 'cogview4'
     elif "Sana" in name:
         model_type = 'sana'
+    elif "HiDream" in name:
+        model_type = 'h1'
     # video models
     elif "CogVideo" in name:
         model_type = 'cogvideo'

modules/sd_detect.py

@@ -90,6 +90,8 @@ def detect_pipeline(f: str, op: str = 'model', warning=True, quiet=False):
                 pipeline = 'custom'
             if 'sd3' in f.lower():
                 guess = 'Stable Diffusion 3'
+            if 'hidream' in f.lower():
+                guess = 'HiDream'
             if 'flux' in f.lower() or 'flex.1' in f.lower():
                 guess = 'FLUX'
                 if size > 11000 and size < 16000:

modules/sd_models.py

@@ -326,6 +326,9 @@ def load_diffuser_force(model_type, checkpoint_info, diffusers_load_config, op='
         elif model_type in ['OmniGen']: # forced pipeline
             from modules.model_omnigen import load_omnigen
             sd_model = load_omnigen(checkpoint_info, diffusers_load_config)
+        elif model_type in ['HiDream']:
+            from modules.model_hidream import load_hidream
+            sd_model = load_hidream(checkpoint_info, diffusers_load_config)
     except Exception as e:
         shared.log.error(f'Load {op}: path="{checkpoint_info.path}" {e}')
         if debug_load:

modules/sd_vae_taesd.py

@@ -52,8 +52,10 @@ def get_model(model_type = 'decoder', variant = None):
     global prev_cls, prev_type, prev_model # pylint: disable=global-statement
     from modules import shared
     cls = shared.sd_model_type
-    if cls == 'ldm':
+    if cls == 'ldm': # original backend
         cls = 'sd'
+    if cls == 'h1': # hidream uses flux vae
+        cls = 'f1'
     variant = variant or shared.opts.taesd_variant
     folder = os.path.join(paths.models_path, "TAESD")
     os.makedirs(folder, exist_ok=True)

modules/shared.py

@@ -403,16 +403,20 @@ options_templates.update(options_section(('sd', "Models & Loading"), {
     "diffusers_offload_max_cpu_memory": OptionInfo(0.90, "Balanced offload CPU high watermark", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.01, "visible": False }),
 
     "advanced_sep": OptionInfo("<h2>Advanced Options</h2>", "", gr.HTML),
-    "sd_checkpoint_autoload": OptionInfo(True, "Model autoload on start"),
+    "sd_checkpoint_autoload": OptionInfo(True, "Model auto-load on start"),
     "sd_checkpoint_autodownload": OptionInfo(True, "Model auto-download on demand"),
     "stream_load": OptionInfo(False, "Model load using streams", gr.Checkbox),
     "diffusers_eval": OptionInfo(True, "Force model eval", gr.Checkbox, {"visible": False }),
-    "diffusers_to_gpu": OptionInfo(False, "Model Load model direct to GPU"),
+    "diffusers_to_gpu": OptionInfo(False, "Model load model direct to GPU"),
     "disable_accelerate": OptionInfo(False, "Disable accelerate", gr.Checkbox, {"visible": False }),
     "sd_model_dict": OptionInfo('None', "Use separate base dict", gr.Dropdown, lambda: {"choices": ['None'] + list_checkpoint_titles(), "visible": False}, refresh=refresh_checkpoints),
     "sd_checkpoint_cache": OptionInfo(0, "Cached models", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1, "visible": not native }),
 }))
 
+options_templates.update(options_section(('model_options', "Models Options"), {
+    "model_sd3_disable_te5": OptionInfo(False, "SD3 disable T5 encoder"),
+}))
+
 options_templates.update(options_section(('vae_encoder', "Variable Auto Encoder"), {
     "sd_vae": OptionInfo("Automatic", "VAE model", gr.Dropdown, lambda: {"choices": shared_items.sd_vae_items()}, refresh=shared_items.refresh_vae_list),
     "diffusers_vae_upcast": OptionInfo("default", "VAE upcasting", gr.Radio, {"choices": ['default', 'true', 'false']}),
@@ -429,7 +433,7 @@ options_templates.update(options_section(('vae_encoder', "Variable Auto Encoder"
 }))
 
 options_templates.update(options_section(('text_encoder', "Text Encoder"), {
-    "sd_text_encoder": OptionInfo('Default', "Text encoder model", gr.Dropdown, lambda: {"choices": shared_items.sd_te_items()}, refresh=shared_items.refresh_te_list),
+    "sd_text_encoder": OptionInfo('Default', "Text encoder model", DropdownEditable, lambda: {"choices": shared_items.sd_te_items()}, refresh=shared_items.refresh_te_list),
     "prompt_attention": OptionInfo("native", "Prompt attention parser", gr.Radio, {"choices": ["native", "compel", "xhinker", "a1111", "fixed"] }),
     "prompt_mean_norm": OptionInfo(False, "Prompt attention normalization", gr.Checkbox),
     "sd_textencoder_cache": OptionInfo(True, "Cache text encoder results", gr.Checkbox, {"visible": False}),

modules/shared_items.py

@@ -32,6 +32,7 @@ pipelines = {
     'CogView 4': getattr(diffusers, 'CogView4Pipeline', None),
     'UniDiffuser': getattr(diffusers, 'UniDiffuserPipeline', None),
     'Amused': getattr(diffusers, 'AmusedPipeline', None),
+    'HiDream': getattr(diffusers, 'HiDreamImagePipeline', None),
 
     # dynamically imported and redefined later
     'Meissonic': getattr(diffusers, 'DiffusionPipeline', None), # dynamically redefined and loaded in sd_models.load_diffuser
@@ -78,7 +79,7 @@ def refresh_unet_list():
 
 def sd_te_items():
     import modules.model_te
-    predefined = ['None', 'T5 FP4', 'T5 FP8', 'T5 INT8', 'T5 QINT8', 'T5 FP16']
+    predefined = ['None']
     return predefined + list(modules.model_te.te_dict)
 
 

scripts/allegrovideo.py

@@ -21,7 +21,7 @@ def hijack_decode(*args, **kwargs):
 
 def hijack_encode_prompt(*args, **kwargs):
     t0 = time.time()
-    res = shared.sd_model.vae.orig_encode_prompt(*args, **kwargs)
+    res = shared.sd_model.orig_encode_prompt(*args, **kwargs)
     t1 = time.time()
     timer.process.add('te', t1-t0)
     shared.log.debug(f'Video: te={shared.sd_model.text_encoder.__class__.__name__} time={t1-t0:.2f}')
@@ -92,7 +92,7 @@ class Script(scripts.Script):
             shared.sd_model.sd_checkpoint_info = sd_checkpoint.CheckpointInfo(repo_id)
             shared.sd_model.sd_model_hash = None
             shared.sd_model.vae.orig_decode = shared.sd_model.vae.decode
-            shared.sd_model.vae.orig_encode_prompt = shared.sd_model.encode_prompt
+            shared.sd_model.orig_encode_prompt = shared.sd_model.encode_prompt
             shared.sd_model.vae.decode = hijack_decode
             shared.sd_model.encode_prompt = hijack_encode_prompt
             shared.sd_model.vae.enable_tiling()

scripts/hunyuanvideo.py

@@ -50,7 +50,7 @@ def hijack_decode(*args, **kwargs):
 
 def hijack_encode_prompt(*args, **kwargs):
     t0 = time.time()
-    res = shared.sd_model.vae.orig_encode_prompt(*args, **kwargs)
+    res = shared.sd_model.orig_encode_prompt(*args, **kwargs)
     t1 = time.time()
     timer.process.add('te', t1-t0)
     shared.log.debug(f'Video: te={shared.sd_model.text_encoder.__class__.__name__} time={t1-t0:.2f}')
@@ -133,7 +133,7 @@ class Script(scripts.Script):
             shared.sd_model.sd_checkpoint_info = sd_checkpoint.CheckpointInfo(models.get(model)['repo'])
             shared.sd_model.sd_model_hash = None
             shared.sd_model.vae.orig_decode = shared.sd_model.vae.decode
-            shared.sd_model.vae.orig_encode_prompt = shared.sd_model.encode_prompt
+            shared.sd_model.orig_encode_prompt = shared.sd_model.encode_prompt
             shared.sd_model.vae.decode = hijack_decode
             shared.sd_model.encode_prompt = hijack_encode_prompt
             shared.sd_model.vae.enable_slicing()

scripts/legacy_allegrovideo.py

@@ -21,7 +21,7 @@ def hijack_decode(*args, **kwargs):
 
 def hijack_encode_prompt(*args, **kwargs):
     t0 = time.time()
-    res = shared.sd_model.vae.orig_encode_prompt(*args, **kwargs)
+    res = shared.sd_model.orig_encode_prompt(*args, **kwargs)
     t1 = time.time()
     timer.process.add('te', t1-t0)
     shared.log.debug(f'Video: te={shared.sd_model.text_encoder.__class__.__name__} time={t1-t0:.2f}')
@@ -92,7 +92,7 @@ class Script(scripts.Script):
             shared.sd_model.sd_checkpoint_info = sd_checkpoint.CheckpointInfo(repo_id)
             shared.sd_model.sd_model_hash = None
             shared.sd_model.vae.orig_decode = shared.sd_model.vae.decode
-            shared.sd_model.vae.orig_encode_prompt = shared.sd_model.encode_prompt
+            shared.sd_model.orig_encode_prompt = shared.sd_model.encode_prompt
             shared.sd_model.vae.decode = hijack_decode
             shared.sd_model.encode_prompt = hijack_encode_prompt
             shared.sd_model.vae.enable_tiling()

scripts/ltxvideo.py

@@ -41,7 +41,7 @@ def hijack_decode(*args, **kwargs):
 
 def hijack_encode_prompt(*args, **kwargs):
     t0 = time.time()
-    res = shared.sd_model.vae.orig_encode_prompt(*args, **kwargs)
+    res = shared.sd_model.orig_encode_prompt(*args, **kwargs)
     t1 = time.time()
     timer.process.add('te', t1-t0)
     shared.log.debug(f'Video: te={shared.sd_model.text_encoder.__class__.__name__} time={t1-t0:.2f}')
@@ -129,7 +129,7 @@ class Script(scripts.Script):
                 )
             sd_models.set_diffuser_options(shared.sd_model)
             shared.sd_model.vae.orig_decode = shared.sd_model.vae.decode
-            shared.sd_model.vae.orig_encode_prompt = shared.sd_model.encode_prompt
+            shared.sd_model.orig_encode_prompt = shared.sd_model.encode_prompt
             shared.sd_model.vae.decode = hijack_decode
             shared.sd_model.encode_prompt = hijack_encode_prompt
             shared.sd_model.sd_checkpoint_info = sd_checkpoint.CheckpointInfo(repo_id)

webui.py

@@ -156,7 +156,7 @@ def initialize():
 
 def load_model():
     if not shared.opts.sd_checkpoint_autoload and shared.cmd_opts.ckpt is None:
-        log.info('Model: auto load disabled')
+        log.info('Model: autoload=False')
     else:
         shared.state.begin('Load')
         thread_model = Thread(target=lambda: shared.sd_model)

wiki

@@ -1 +1 @@
-Subproject commit f4b7d2e573185f08ef6e06a6a97763aa83d59701
+Subproject commit 9f5ee99a8e6c437801a8727f73acfffe2f0b926e