diff --git a/CHANGELOG.md b/CHANGELOG.md index 00e3bef4b..ccacfd211 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,13 @@ # Change Log for SD.Next +## Update for 2026-04-02 + +- **Models** + - [AiArtLab SDXS-1B](https://huggingface.co/AiArtLab/sdxs-1b) Simple Diffusion XS *(training still in progress)* + this model combines Qwen3.5-1.8B text encoder with SDXL-style UNET with only 1.6B parameters and custom 32ch VAE +- **Internal** + - additional typing and typechecks, thanks @awsr + ## Update for 2026-04-01 ### Highlights for 2026-04-01 diff --git a/data/reference.json b/data/reference.json index d2ea919c7..87c8c3d37 100644 --- a/data/reference.json +++ b/data/reference.json @@ -864,6 +864,16 @@ "extras": "sampler: Default, cfg_scale: 1.5, steps: 50", "size": 15.3, "date": "2025 January" + }, + + "AiArtLab SDXS-1B": { + "path": "AiArtLab/sdxs-1b", + "preview": "AiArtLab--sdxs-1b.jpg", + "desc": "Simple Diffusion XS (train in progress) combines Qwen3.5-1.8B text encoder with SDXL-style UNET with only 1.6B parameters and custom 32ch VAE", + "skip": true, + "extras": "sampler: Default", + "size": 15.3, + "date": "2026 January" } } diff --git a/models/Reference/AiArtLab--sdxs-1b.jpg b/models/Reference/AiArtLab--sdxs-1b.jpg new file mode 100644 index 000000000..e69de29bb diff --git a/modules/modeldata.py b/modules/modeldata.py index fa0d94f64..5c38e1820 100644 --- a/modules/modeldata.py +++ b/modules/modeldata.py @@ -100,6 +100,8 @@ def get_model_type(pipe): model_type = 'hunyuanimage3' elif 'HunyuanImage' in name: model_type = 'hunyuanimage' + elif 'sdxs-1b' in name: + model_type = 'sdxs' # video models elif "Kandinsky5" in name and '2V' in name: model_type = 'kandinsky5video' diff --git a/modules/processing_class.py b/modules/processing_class.py index b0e3df0ec..299f9f9b4 100644 --- a/modules/processing_class.py +++ b/modules/processing_class.py @@ -325,8 +325,8 @@ class StableDiffusionProcessing: # initializers self.prompt = prompt - self.seed = seed - self.subseed = subseed + self.seed = int(seed) + self.subseed = int(subseed) self.subseed_strength = subseed_strength self.seed_resize_from_h = seed_resize_from_h self.seed_resize_from_w = seed_resize_from_w diff --git a/modules/sd_detect.py b/modules/sd_detect.py index 0ba0a8ce4..3ee2f8428 100644 --- a/modules/sd_detect.py +++ b/modules/sd_detect.py @@ -150,6 +150,8 @@ def guess_by_name(fn, current_guess): new_guess = 'Ovis-Image' elif 'glm-image' in fn.lower(): new_guess = 'GLM-Image' + elif 'sdxs-1b' in fn.lower(): + new_guess = 'SDXS' if debug_load: log.trace(f'Autodetect: method=name file="{fn}" previous="{current_guess}" current="{new_guess}"') return new_guess or current_guess @@ -166,6 +168,8 @@ def guess_by_diffusers(fn, current_guess): if name is not None and name in exclude_by_name: return current_guess, None cls = index.get('_class_name', None) + if isinstance(cls, list): + cls = cls[-1] if cls is not None: pipeline = getattr(diffusers, cls, None) if pipeline is None: diff --git a/modules/sd_hijack_te.py b/modules/sd_hijack_te.py index c853b6417..b35746cfa 100644 --- a/modules/sd_hijack_te.py +++ b/modules/sd_hijack_te.py @@ -8,7 +8,7 @@ def hijack_encode_prompt(*args, **kwargs): jobid = shared.state.begin('TE Encode') t0 = time.time() if 'max_sequence_length' in kwargs and kwargs['max_sequence_length'] is not None: - kwargs['max_sequence_length'] = max(kwargs['max_sequence_length'], os.environ.get('HIDREAM_MAX_SEQUENCE_LENGTH', 256)) + kwargs['max_sequence_length'] = max(kwargs['max_sequence_length'], os.environ.get('MAX_SEQUENCE_LENGTH', 256)) try: prompt = kwargs.get('prompt', None) or (args[0] if len(args) > 0 else None) if prompt is not None: @@ -20,8 +20,6 @@ def hijack_encode_prompt(*args, **kwargs): res = None t1 = time.time() timer.process.add('te', t1-t0) - # if hasattr(shared.sd_model, "maybe_free_model_hooks"): - # shared.sd_model.maybe_free_model_hooks() shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model) shared.state.end(jobid) return res diff --git a/modules/sd_models.py b/modules/sd_models.py index a96d62c9d..d33396b9b 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -507,6 +507,10 @@ def load_diffuser_force(detected_model_type, checkpoint_info, diffusers_load_con from pipelines.model_glm import load_glm_image sd_model = load_glm_image(checkpoint_info, diffusers_load_config) allow_post_quant = False + elif model_type in ['SDXS']: + from pipelines.model_sdxs import load_sdxs + sd_model = load_sdxs(checkpoint_info, diffusers_load_config) + allow_post_quant = False except Exception as e: log.error(f'Load {op}: path="{checkpoint_info.path}" {e}') # if debug_load: diff --git a/modules/shared_items.py b/modules/shared_items.py index b787cd20c..e33258d8f 100644 --- a/modules/shared_items.py +++ b/modules/shared_items.py @@ -65,6 +65,7 @@ pipelines = { 'HunyuanImage3': getattr(diffusers, 'DiffusionPipeline', None), 'ChronoEdit': getattr(diffusers, 'DiffusionPipeline', None), 'Anima': getattr(diffusers, 'DiffusionPipeline', None), + 'SDXS': getattr(diffusers, 'DiffusionPipeline', None), } diff --git a/pipelines/model_sdxs.py b/pipelines/model_sdxs.py new file mode 100644 index 000000000..f4cc142ef --- /dev/null +++ b/pipelines/model_sdxs.py @@ -0,0 +1,59 @@ +import time +import diffusers +import transformers +from modules import shared, devices, errors, timer, sd_models, model_quant, sd_hijack_vae +from modules.logger import log +from pipelines import generic + + +def hijack_encode_text(prompt: str | list[str]): + jobid = shared.state.begin('TE Encode') + t0 = time.time() + try: + prompt = shared.sd_model.refine_prompts(prompt) + except Exception as e: + log.error(f'Encode prompt: {e}') + errors.display(e, 'Encode prompt') + try: + res = shared.sd_model.orig_encode_text(prompt) + except Exception as e: + log.error(f'Encode prompt: {e}') + errors.display(e, 'Encode prompt') + res = None + t1 = time.time() + timer.process.add('te', t1-t0) + shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model) + shared.state.end(jobid) + return res + + +def load_sdxs(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=SDXS repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}') + + text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3_5ForConditionalGeneration, load_config=diffusers_load_config, allow_shared=False) + + pipe = diffusers.DiffusionPipeline.from_pretrained( + repo_id, + text_encoder=text_encoder, + cache_dir=shared.opts.diffusers_dir, + trust_remote_code=True, + **load_args, + ) + pipe.task_args = { + 'generator': None, + 'output_type': 'np', + } + + pipe.orig_encode_text = pipe.encode_text + pipe.encode_text = hijack_encode_text + sd_hijack_vae.init_hijack(pipe) + del text_encoder + + devices.torch_gc(force=True, reason='load') + return pipe diff --git a/wiki b/wiki index d54ade8e5..a9b73a500 160000 --- a/wiki +++ b/wiki @@ -1 +1 @@ -Subproject commit d54ade8e5f79a62b5228de6406400c8eda71b67f +Subproject commit a9b73a50018a08a1ac35fb15c7f7c86515b79d13