diff options
Diffstat (limited to 'modules')
| -rw-r--r-- | modules/lowvram.py | 10 | ||||
| -rw-r--r-- | modules/processing.py | 41 | ||||
| -rw-r--r-- | modules/sd_models.py | 5 | ||||
| -rw-r--r-- | modules/sd_models_config.py | 7 | ||||
| -rw-r--r-- | modules/sd_samplers_compvis.py | 31 | ||||
| -rw-r--r-- | modules/sd_samplers_kdiffusion.py | 19 |
6 files changed, 84 insertions, 29 deletions
diff --git a/modules/lowvram.py b/modules/lowvram.py index 042a0254..e254cc13 100644 --- a/modules/lowvram.py +++ b/modules/lowvram.py @@ -55,12 +55,12 @@ def setup_for_low_vram(sd_model, use_medvram): if hasattr(sd_model.cond_stage_model, 'model'):
sd_model.cond_stage_model.transformer = sd_model.cond_stage_model.model
- # remove four big modules, cond, first_stage, depth (if applicable), and unet from the model and then
+ # remove several big modules: cond, first_stage, depth/embedder (if applicable), and unet from the model and then
# send the model to GPU. Then put modules back. the modules will be in CPU.
- stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, getattr(sd_model, 'depth_model', None), sd_model.model
- sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.model = None, None, None, None
+ stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, getattr(sd_model, 'depth_model', None), getattr(sd_model, 'embedder', None), sd_model.model
+ sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.embedder, sd_model.model = None, None, None, None, None
sd_model.to(devices.device)
- sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.model = stored
+ sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.embedder, sd_model.model = stored
# register hooks for those the first three models
sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu)
@@ -69,6 +69,8 @@ def setup_for_low_vram(sd_model, use_medvram): sd_model.first_stage_model.decode = first_stage_model_decode_wrap
if sd_model.depth_model:
sd_model.depth_model.register_forward_pre_hook(send_me_to_gpu)
+ if sd_model.embedder:
+ sd_model.embedder.register_forward_pre_hook(send_me_to_gpu)
parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
if hasattr(sd_model.cond_stage_model, 'model'):
diff --git a/modules/processing.py b/modules/processing.py index 59717b4c..1451811c 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -78,21 +78,27 @@ def apply_overlay(image, paste_loc, index, overlays): def txt2img_image_conditioning(sd_model, x, width, height):
- if sd_model.model.conditioning_key not in {'hybrid', 'concat'}:
- # Dummy zero conditioning if we're not using inpainting model.
- # Still takes up a bit of memory, but no encoder call.
- # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size.
- return x.new_zeros(x.shape[0], 5, 1, 1, dtype=x.dtype, device=x.device)
+ if sd_model.model.conditioning_key in {'hybrid', 'concat'}: # Inpainting models
- # The "masked-image" in this case will just be all zeros since the entire image is masked.
- image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
- image_conditioning = sd_model.get_first_stage_encoding(sd_model.encode_first_stage(image_conditioning))
+ # The "masked-image" in this case will just be all zeros since the entire image is masked.
+ image_conditioning = torch.zeros(x.shape[0], 3, height, width, device=x.device)
+ image_conditioning = sd_model.get_first_stage_encoding(sd_model.encode_first_stage(image_conditioning))
- # Add the fake full 1s mask to the first dimension.
- image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
- image_conditioning = image_conditioning.to(x.dtype)
+ # Add the fake full 1s mask to the first dimension.
+ image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
+ image_conditioning = image_conditioning.to(x.dtype)
- return image_conditioning
+ return image_conditioning
+
+ elif sd_model.model.conditioning_key == "crossattn-adm": # UnCLIP models
+
+ return x.new_zeros(x.shape[0], 2*sd_model.noise_augmentor.time_embed.dim, dtype=x.dtype, device=x.device)
+
+ else:
+ # Dummy zero conditioning if we're not using inpainting or unclip models.
+ # Still takes up a bit of memory, but no encoder call.
+ # Pretty sure we can just make this a 1x1 image since its not going to be used besides its batch size.
+ return x.new_zeros(x.shape[0], 5, 1, 1, dtype=x.dtype, device=x.device)
class StableDiffusionProcessing:
@@ -190,6 +196,14 @@ class StableDiffusionProcessing: return conditioning_image
+ def unclip_image_conditioning(self, source_image):
+ c_adm = self.sd_model.embedder(source_image)
+ if self.sd_model.noise_augmentor is not None:
+ noise_level = 0 # TODO: Allow other noise levels?
+ c_adm, noise_level_emb = self.sd_model.noise_augmentor(c_adm, noise_level=repeat(torch.tensor([noise_level]).to(c_adm.device), '1 -> b', b=c_adm.shape[0]))
+ c_adm = torch.cat((c_adm, noise_level_emb), 1)
+ return c_adm
+
def inpainting_image_conditioning(self, source_image, latent_image, image_mask=None):
self.is_using_inpainting_conditioning = True
@@ -241,6 +255,9 @@ class StableDiffusionProcessing: if self.sampler.conditioning_key in {'hybrid', 'concat'}:
return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
+ if self.sampler.conditioning_key == "crossattn-adm":
+ return self.unclip_image_conditioning(source_image)
+
# Dummy zero conditioning if we're not using inpainting or depth model.
return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1)
diff --git a/modules/sd_models.py b/modules/sd_models.py index f0cb1240..c1a80d82 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -383,6 +383,11 @@ def repair_config(sd_config): elif shared.cmd_opts.upcast_sampling:
sd_config.model.params.unet_config.params.use_fp16 = True
+ # For UnCLIP-L, override the hardcoded karlo directory
+ if hasattr(sd_config.model.params, "noise_aug_config") and hasattr(sd_config.model.params.noise_aug_config.params, "clip_stats_path"):
+ karlo_path = os.path.join(paths.models_path, 'karlo')
+ sd_config.model.params.noise_aug_config.params.clip_stats_path = sd_config.model.params.noise_aug_config.params.clip_stats_path.replace("checkpoints/karlo_models", karlo_path)
+
sd1_clip_weight = 'cond_stage_model.transformer.text_model.embeddings.token_embedding.weight'
sd2_clip_weight = 'cond_stage_model.model.transformer.resblocks.0.attn.in_proj_weight'
diff --git a/modules/sd_models_config.py b/modules/sd_models_config.py index 91c21700..9398f528 100644 --- a/modules/sd_models_config.py +++ b/modules/sd_models_config.py @@ -14,6 +14,8 @@ config_sd2 = os.path.join(sd_repo_configs_path, "v2-inference.yaml") config_sd2v = os.path.join(sd_repo_configs_path, "v2-inference-v.yaml")
config_sd2_inpainting = os.path.join(sd_repo_configs_path, "v2-inpainting-inference.yaml")
config_depth_model = os.path.join(sd_repo_configs_path, "v2-midas-inference.yaml")
+config_unclip = os.path.join(sd_repo_configs_path, "v2-1-stable-unclip-l-inference.yaml")
+config_unopenclip = os.path.join(sd_repo_configs_path, "v2-1-stable-unclip-h-inference.yaml")
config_inpainting = os.path.join(sd_configs_path, "v1-inpainting-inference.yaml")
config_instruct_pix2pix = os.path.join(sd_configs_path, "instruct-pix2pix.yaml")
config_alt_diffusion = os.path.join(sd_configs_path, "alt-diffusion-inference.yaml")
@@ -65,9 +67,14 @@ def is_using_v_parameterization_for_sd2(state_dict): def guess_model_config_from_state_dict(sd, filename):
sd2_cond_proj_weight = sd.get('cond_stage_model.model.transformer.resblocks.0.attn.in_proj_weight', None)
diffusion_model_input = sd.get('model.diffusion_model.input_blocks.0.0.weight', None)
+ sd2_variations_weight = sd.get('embedder.model.ln_final.weight', None)
if sd.get('depth_model.model.pretrained.act_postprocess3.0.project.0.bias', None) is not None:
return config_depth_model
+ elif sd2_variations_weight is not None and sd2_variations_weight.shape[0] == 768:
+ return config_unclip
+ elif sd2_variations_weight is not None and sd2_variations_weight.shape[0] == 1024:
+ return config_unopenclip
if sd2_cond_proj_weight is not None and sd2_cond_proj_weight.shape[1] == 1024:
if diffusion_model_input.shape[1] == 9:
diff --git a/modules/sd_samplers_compvis.py b/modules/sd_samplers_compvis.py index 083da18c..bfcc5574 100644 --- a/modules/sd_samplers_compvis.py +++ b/modules/sd_samplers_compvis.py @@ -70,8 +70,13 @@ class VanillaStableDiffusionSampler: # Have to unwrap the inpainting conditioning here to perform pre-processing
image_conditioning = None
+ uc_image_conditioning = None
if isinstance(cond, dict):
- image_conditioning = cond["c_concat"][0]
+ if self.conditioning_key == "crossattn-adm":
+ image_conditioning = cond["c_adm"]
+ uc_image_conditioning = unconditional_conditioning["c_adm"]
+ else:
+ image_conditioning = cond["c_concat"][0]
cond = cond["c_crossattn"][0]
unconditional_conditioning = unconditional_conditioning["c_crossattn"][0]
@@ -98,8 +103,12 @@ class VanillaStableDiffusionSampler: # 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.
if image_conditioning is not None:
- cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]}
- unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
+ if self.conditioning_key == "crossattn-adm":
+ cond = {"c_adm": image_conditioning, "c_crossattn": [cond]}
+ unconditional_conditioning = {"c_adm": uc_image_conditioning, "c_crossattn": [unconditional_conditioning]}
+ else:
+ cond = {"c_concat": [image_conditioning], "c_crossattn": [cond]}
+ unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
return x, ts, cond, unconditional_conditioning
@@ -176,8 +185,12 @@ class VanillaStableDiffusionSampler: # Wrap the conditioning models with additional image conditioning for inpainting model
if image_conditioning is not None:
- conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]}
- unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
+ if self.conditioning_key == "crossattn-adm":
+ conditioning = {"c_adm": image_conditioning, "c_crossattn": [conditioning]}
+ unconditional_conditioning = {"c_adm": torch.zeros_like(image_conditioning), "c_crossattn": [unconditional_conditioning]}
+ else:
+ conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]}
+ unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
samples = self.launch_sampling(t_enc + 1, lambda: self.sampler.decode(x1, conditioning, t_enc, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning))
@@ -195,8 +208,12 @@ class VanillaStableDiffusionSampler: # Wrap the conditioning models with additional image conditioning for inpainting model
# dummy_for_plms is needed because PLMS code checks the first item in the dict to have the right shape
if image_conditioning is not None:
- conditioning = {"dummy_for_plms": np.zeros((conditioning.shape[0],)), "c_crossattn": [conditioning], "c_concat": [image_conditioning]}
- unconditional_conditioning = {"c_crossattn": [unconditional_conditioning], "c_concat": [image_conditioning]}
+ if self.conditioning_key == "crossattn-adm":
+ conditioning = {"dummy_for_plms": np.zeros((conditioning.shape[0],)), "c_crossattn": [conditioning], "c_adm": image_conditioning}
+ unconditional_conditioning = {"c_crossattn": [unconditional_conditioning], "c_adm": torch.zeros_like(image_conditioning)}
+ else:
+ conditioning = {"dummy_for_plms": np.zeros((conditioning.shape[0],)), "c_crossattn": [conditioning], "c_concat": [image_conditioning]}
+ unconditional_conditioning = {"c_crossattn": [unconditional_conditioning], "c_concat": [image_conditioning]}
samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0])
diff --git a/modules/sd_samplers_kdiffusion.py b/modules/sd_samplers_kdiffusion.py index 93f0e55a..e9f08518 100644 --- a/modules/sd_samplers_kdiffusion.py +++ b/modules/sd_samplers_kdiffusion.py @@ -92,14 +92,21 @@ class CFGDenoiser(torch.nn.Module): batch_size = len(conds_list)
repeats = [len(conds_list[i]) for i in range(batch_size)]
+ if shared.sd_model.model.conditioning_key == "crossattn-adm":
+ image_uncond = torch.zeros_like(image_cond)
+ make_condition_dict = lambda c_crossattn, c_adm: {"c_crossattn": c_crossattn, "c_adm": c_adm}
+ else:
+ image_uncond = image_cond
+ make_condition_dict = lambda c_crossattn, c_concat: {"c_crossattn": c_crossattn, "c_concat": [c_concat]}
+
if not is_edit_model:
x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x])
sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma])
- image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_cond])
+ image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_uncond])
else:
x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x] + [x])
sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma] + [sigma])
- image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_cond] + [torch.zeros_like(self.init_latent)])
+ image_cond_in = torch.cat([torch.stack([image_cond[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [image_uncond] + [torch.zeros_like(self.init_latent)])
denoiser_params = CFGDenoiserParams(x_in, image_cond_in, sigma_in, state.sampling_step, state.sampling_steps, tensor, uncond)
cfg_denoiser_callback(denoiser_params)
@@ -116,13 +123,13 @@ class CFGDenoiser(torch.nn.Module): cond_in = torch.cat([tensor, uncond, uncond])
if shared.batch_cond_uncond:
- x_out = self.inner_model(x_in, sigma_in, cond={"c_crossattn": [cond_in], "c_concat": [image_cond_in]})
+ x_out = self.inner_model(x_in, sigma_in, cond=make_condition_dict([cond_in], image_cond_in))
else:
x_out = torch.zeros_like(x_in)
for batch_offset in range(0, x_out.shape[0], batch_size):
a = batch_offset
b = a + batch_size
- x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": [cond_in[a:b]], "c_concat": [image_cond_in[a:b]]})
+ x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=make_condition_dict([cond_in[a:b]], image_cond_in[a:b]))
else:
x_out = torch.zeros_like(x_in)
batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size
@@ -135,9 +142,9 @@ class CFGDenoiser(torch.nn.Module): else:
c_crossattn = torch.cat([tensor[a:b]], uncond)
- x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond={"c_crossattn": c_crossattn, "c_concat": [image_cond_in[a:b]]})
+ x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=make_condition_dict(c_crossattn, image_cond_in[a:b]))
- x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond={"c_crossattn": [uncond], "c_concat": [image_cond_in[-uncond.shape[0]:]]})
+ x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=make_condition_dict([uncond], image_cond_in[-uncond.shape[0]:]))
denoised_params = CFGDenoisedParams(x_out, state.sampling_step, state.sampling_steps)
cfg_denoised_callback(denoised_params)
|