Merge pull request #7710 from space-nuko/unipc

Implement UniPC sampler

javascript/hints.js

@@ -6,6 +6,7 @@ titles = {
 	"GFPGAN": "Restore low quality faces using GFPGAN neural network",
 	"Euler a": "Euler Ancestral - very creative, each can get a completely different picture depending on step count, setting steps higher than 30-40 does not help",
 	"DDIM": "Denoising Diffusion Implicit Models - best at inpainting",
+	"UniPC": "Unified Predictor-Corrector Framework for Fast Sampling of Diffusion Models",
 	"DPM adaptive": "Ignores step count - uses a number of steps determined by the CFG and resolution", 
 
 	"Batch count": "How many batches of images to create (has no impact on generation performance or VRAM usage)",

modules/models/diffusion/uni_pc/__init__.py

+from .sampler import UniPCSampler

modules/models/diffusion/uni_pc/sampler.py

+"""SAMPLING ONLY."""
+
+import torch
+
+from .uni_pc import NoiseScheduleVP, model_wrapper, UniPC
+from modules import shared
+
+class UniPCSampler(object):
+    def __init__(self, model, **kwargs):
+        super().__init__()
+        self.model = model
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device)
+        self.before_sample = None
+        self.after_sample = None
+        self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod))
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def set_hooks(self, before_sample, after_sample, after_update):
+        self.before_sample = before_sample
+        self.after_sample = after_sample
+        self.after_update = after_update
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                ctmp = conditioning[list(conditioning.keys())[0]]
+                while isinstance(ctmp, list): ctmp = ctmp[0]
+                cbs = ctmp.shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+            elif isinstance(conditioning, list):
+                for ctmp in conditioning:
+                    if ctmp.shape[0] != batch_size:
+                        print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for UniPC sampling is {size}')
+
+        device = self.model.betas.device
+        if x_T is None:
+            img = torch.randn(size, device=device)
+        else:
+            img = x_T
+
+        ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod)
+
+        # SD 1.X is "noise", SD 2.X is "v"
+        model_type = "v" if self.model.parameterization == "v" else "noise"
+
+        model_fn = model_wrapper(
+            lambda x, t, c: self.model.apply_model(x, t, c),
+            ns,
+            model_type=model_type,
+            guidance_type="classifier-free",
+            #condition=conditioning,
+            #unconditional_condition=unconditional_conditioning,
+            guidance_scale=unconditional_guidance_scale,
+        )
+
+        uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False, variant=shared.opts.uni_pc_variant, condition=conditioning, unconditional_condition=unconditional_conditioning, before_sample=self.before_sample, after_sample=self.after_sample, after_update=self.after_update)
+        x = uni_pc.sample(img, steps=S, skip_type=shared.opts.uni_pc_skip_type, method="multistep", order=shared.opts.uni_pc_order, lower_order_final=shared.opts.uni_pc_lower_order_final)
+
+        return x.to(device), None

modules/processing.py

@@ -888,7 +888,9 @@ class StableDiffusionProcessingTxt2Img(StableDiffusionProcessing):
 
         shared.state.nextjob()
 
-        img2img_sampler_name = self.sampler_name if self.sampler_name != 'PLMS' else 'DDIM'  # PLMS does not support img2img so we just silently switch ot DDIM
+        img2img_sampler_name = self.sampler_name
+        if self.sampler_name in ['PLMS', 'UniPC']:  # PLMS/UniPC do not support img2img so we just silently switch to DDIM
+            img2img_sampler_name = 'DDIM'
         self.sampler = sd_samplers.create_sampler(img2img_sampler_name, self.sd_model)
 
         samples = samples[:, :, self.truncate_y//2:samples.shape[2]-(self.truncate_y+1)//2, self.truncate_x//2:samples.shape[3]-(self.truncate_x+1)//2]

modules/sd_samplers.py

@@ -32,7 +32,7 @@ def set_samplers():
     global samplers, samplers_for_img2img
 
     hidden = set(shared.opts.hide_samplers)
-    hidden_img2img = set(shared.opts.hide_samplers + ['PLMS'])
+    hidden_img2img = set(shared.opts.hide_samplers + ['PLMS', 'UniPC'])
 
     samplers = [x for x in all_samplers if x.name not in hidden]
     samplers_for_img2img = [x for x in all_samplers if x.name not in hidden_img2img]

modules/sd_samplers_compvis.py

@@ -7,19 +7,27 @@ import torch
 
 from modules.shared import state
 from modules import sd_samplers_common, prompt_parser, shared
+import modules.models.diffusion.uni_pc
 
 
 samplers_data_compvis = [
     sd_samplers_common.SamplerData('DDIM', lambda model: VanillaStableDiffusionSampler(ldm.models.diffusion.ddim.DDIMSampler, model), [], {}),
     sd_samplers_common.SamplerData('PLMS', lambda model: VanillaStableDiffusionSampler(ldm.models.diffusion.plms.PLMSSampler, model), [], {}),
+    sd_samplers_common.SamplerData('UniPC', lambda model: VanillaStableDiffusionSampler(modules.models.diffusion.uni_pc.UniPCSampler, model), [], {}),
 ]
 
 
 class VanillaStableDiffusionSampler:
     def __init__(self, constructor, sd_model):
         self.sampler = constructor(sd_model)
+        self.is_ddim = hasattr(self.sampler, 'p_sample_ddim')
         self.is_plms = hasattr(self.sampler, 'p_sample_plms')
-        self.orig_p_sample_ddim = self.sampler.p_sample_plms if self.is_plms else self.sampler.p_sample_ddim
+        self.is_unipc = isinstance(self.sampler, modules.models.diffusion.uni_pc.UniPCSampler)
+        self.orig_p_sample_ddim = None
+        if self.is_plms:
+            self.orig_p_sample_ddim = self.sampler.p_sample_plms
+        elif self.is_ddim:
+            self.orig_p_sample_ddim = self.sampler.p_sample_ddim
         self.mask = None
         self.nmask = None
         self.init_latent = None
@@ -45,6 +53,15 @@ class VanillaStableDiffusionSampler:
             return self.last_latent
 
     def p_sample_ddim_hook(self, x_dec, cond, ts, unconditional_conditioning, *args, **kwargs):
+        x_dec, ts, cond, unconditional_conditioning = self.before_sample(x_dec, ts, cond, unconditional_conditioning)
+
+        res = self.orig_p_sample_ddim(x_dec, cond, ts, unconditional_conditioning=unconditional_conditioning, *args, **kwargs)
+
+        x_dec, ts, cond, unconditional_conditioning, res = self.after_sample(x_dec, ts, cond, unconditional_conditioning, res)
+
+        return res
+
+    def before_sample(self, x, ts, cond, unconditional_conditioning):
         if state.interrupted or state.skipped:
             raise sd_samplers_common.InterruptedException
 
@@ -76,7 +93,7 @@ class VanillaStableDiffusionSampler:
 
         if self.mask is not None:
             img_orig = self.sampler.model.q_sample(self.init_latent, ts)
-            x_dec = img_orig * self.mask + self.nmask * x_dec
+            x = img_orig * self.mask + self.nmask * x
 
         # Wrap the image conditioning back up since the DDIM code can accept the dict directly.
         # Note that they need to be lists because it just concatenates them later.
@@ -84,12 +101,13 @@ class VanillaStableDiffusionSampler:
             cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]}
             unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
 
-        res = self.orig_p_sample_ddim(x_dec, cond, ts, unconditional_conditioning=unconditional_conditioning, *args, **kwargs)
+        return x, ts, cond, unconditional_conditioning
 
+    def update_step(self, last_latent):
         if self.mask is not None:
-            self.last_latent = self.init_latent * self.mask + self.nmask * res[1]
+            self.last_latent = self.init_latent * self.mask + self.nmask * last_latent
         else:
-            self.last_latent = res[1]
+            self.last_latent = last_latent
 
         sd_samplers_common.store_latent(self.last_latent)
 
@@ -97,7 +115,14 @@ class VanillaStableDiffusionSampler:
         state.sampling_step = self.step
         shared.total_tqdm.update()
 
-        return res
+    def after_sample(self, x, ts, cond, uncond, res):
+        if not self.is_unipc:
+            self.update_step(res[1])
+
+        return x, ts, cond, uncond, res
+
+    def unipc_after_update(self, x, model_x):
+        self.update_step(x)
 
     def initialize(self, p):
         self.eta = p.eta if p.eta is not None else shared.opts.eta_ddim
@@ -107,16 +132,20 @@ class VanillaStableDiffusionSampler:
         for fieldname in ['p_sample_ddim', 'p_sample_plms']:
             if hasattr(self.sampler, fieldname):
                 setattr(self.sampler, fieldname, self.p_sample_ddim_hook)
+        if self.is_unipc:
+            self.sampler.set_hooks(lambda x, t, c, u: self.before_sample(x, t, c, u), lambda x, t, c, u, r: self.after_sample(x, t, c, u, r), lambda x, mx: self.unipc_after_update(x, mx))
 
         self.mask = p.mask if hasattr(p, 'mask') else None
         self.nmask = p.nmask if hasattr(p, 'nmask') else None
 
     def adjust_steps_if_invalid(self, p, num_steps):
-        if (self.config.name == 'DDIM' and p.ddim_discretize == 'uniform') or (self.config.name == 'PLMS'):
+        if ((self.config.name == 'DDIM') and p.ddim_discretize == 'uniform') or (self.config.name == 'PLMS') or (self.config.name == 'UniPC'):
+            if self.config.name == 'UniPC' and num_steps < shared.opts.uni_pc_order:
+                num_steps = shared.opts.uni_pc_order
             valid_step = 999 / (1000 // num_steps)
             if valid_step == math.floor(valid_step):
                 return int(valid_step) + 1
-        
+
         return num_steps
 
     def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):

modules/shared.py

@@ -485,6 +485,10 @@ options_templates.update(options_section(('sampler-params', "Sampler parameters"
     's_noise': OptionInfo(1.0, "sigma noise", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     'eta_noise_seed_delta': OptionInfo(0, "Eta noise seed delta", gr.Number, {"precision": 0}),
     'always_discard_next_to_last_sigma': OptionInfo(False, "Always discard next-to-last sigma"),
+    'uni_pc_variant': OptionInfo("bh1", "UniPC variant", gr.Radio, {"choices": ["bh1", "bh2", "vary_coeff"]}),
+    'uni_pc_skip_type': OptionInfo("time_uniform", "UniPC skip type", gr.Radio, {"choices": ["time_uniform", "time_quadratic", "logSNR"]}),
+    'uni_pc_order': OptionInfo(3, "UniPC order (must be < sampling steps)", gr.Slider, {"minimum": 1, "maximum": 50, "step": 1}),
+    'uni_pc_lower_order_final': OptionInfo(True, "UniPC lower order final"),
 }))
 
 options_templates.update(options_section(('postprocessing', "Postprocessing"), {

scripts/xyz_grid.py

@@ -128,6 +128,10 @@ def apply_styles(p: StableDiffusionProcessingTxt2Img, x: str, _):
     p.styles.extend(x.split(','))
 
 
+def apply_uni_pc_order(p, x, xs):
+    opts.data["uni_pc_order"] = min(x, p.steps - 1)
+
+
 def format_value_add_label(p, opt, x):
     if type(x) == float:
         x = round(x, 8)
@@ -205,6 +209,7 @@ axis_options = [
     AxisOptionImg2Img("Cond. Image Mask Weight", float, apply_field("inpainting_mask_weight")),
     AxisOption("VAE", str, apply_vae, cost=0.7, choices=lambda: list(sd_vae.vae_dict)),
     AxisOption("Styles", str, apply_styles, choices=lambda: list(shared.prompt_styles.styles)),
+    AxisOption("UniPC Order", int, apply_uni_pc_order, cost=0.5),
 ]
 
 
@@ -316,9 +321,11 @@ class SharedSettingsStackHelper(object):
     def __enter__(self):
         self.CLIP_stop_at_last_layers = opts.CLIP_stop_at_last_layers
         self.vae = opts.sd_vae
+        self.uni_pc_order = opts.uni_pc_order
   
     def __exit__(self, exc_type, exc_value, tb):
         opts.data["sd_vae"] = self.vae
+        opts.data["uni_pc_order"] = self.uni_pc_order
         modules.sd_models.reload_model_weights()
         modules.sd_vae.reload_vae_weights()
 

test/basic_features/txt2img_test.py

@@ -66,6 +66,8 @@ class TestTxt2ImgWorking(unittest.TestCase):
         self.assertEqual(requests.post(self.url_txt2img, json=self.simple_txt2img).status_code, 200)
         self.simple_txt2img["sampler_index"] = "DDIM"
         self.assertEqual(requests.post(self.url_txt2img, json=self.simple_txt2img).status_code, 200)
+        self.simple_txt2img["sampler_index"] = "UniPC"
+        self.assertEqual(requests.post(self.url_txt2img, json=self.simple_txt2img).status_code, 200)
 
     def test_txt2img_multiple_batches_performed(self):
         self.simple_txt2img["n_iter"] = 2