diff --git a/scripts/img2imgalt.py b/scripts/img2imgalt.py index 1e833fa898f..7af93bae974 100644 --- a/scripts/img2imgalt.py +++ b/scripts/img2imgalt.py @@ -11,8 +11,13 @@ import torch import k_diffusion as K -def find_noise_for_image(p, cond, uncond, cfg_scale, steps): - x = p.init_latent +# Debugging notes - the original method apply_model is being called for sd1.5 is in modules.sd_hijack_utils and is ldm.models.diffusion.ddpm.LatentDiffusion +# For sdxl - OpenAIWrapper will be called, which will call the underlying diffusion_model +# When controlnet is enabled, the underlying model is not available to use, therefore we skip + +@torch.no_grad() +def find_noise_for_image(p, cond, uncond, cfg_scale, steps, skip_sdxl_vector): + x = p.init_latent.clone() s_in = x.new_ones([x.shape[0]]) if shared.sd_model.parameterization == "v": @@ -30,41 +35,30 @@ def find_noise_for_image(p, cond, uncond, cfg_scale, steps): x_in = torch.cat([x] * 2) sigma_in = torch.cat([sigmas[i] * s_in] * 2) - cond_in = torch.cat([uncond, cond]) - image_conditioning = torch.cat([p.image_conditioning] * 2) - cond_in = {"c_concat": [image_conditioning], "c_crossattn": [cond_in]} + if shared.sd_model.is_sdxl: + cond_in = {"crossattn": [torch.cat([uncond['crossattn'], cond['crossattn']])], "vector": [torch.cat([uncond['vector'], cond['vector']])]} + else: + cond_in = {"c_concat": [torch.cat([p.image_conditioning] * 2)], "c_crossattn": [torch.cat([uncond, cond])]} - c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]] t = dnw.sigma_to_t(sigma_in) - - eps = shared.sd_model.apply_model(x_in * c_in, t, cond=cond_in) - denoised_uncond, denoised_cond = (x_in + eps * c_out).chunk(2) - - denoised = denoised_uncond + (denoised_cond - denoised_uncond) * cfg_scale - - d = (x - denoised) / sigmas[i] dt = sigmas[i] - sigmas[i - 1] - - x = x + d * dt + x += noise_from_model(x, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip, skip_sdxl_vector) sd_samplers_common.store_latent(x) # This shouldn't be necessary, but solved some VRAM issues - del x_in, sigma_in, cond_in, c_out, c_in, t, - del eps, denoised_uncond, denoised_cond, denoised, d, dt + del x_in, sigma_in, cond_in, t, dt shared.state.nextjob() - return x / x.std() - - -Cached = namedtuple("Cached", ["noise", "cfg_scale", "steps", "latent", "original_prompt", "original_negative_prompt", "sigma_adjustment"]) + return x, sigmas[-1] # Based on changes suggested by briansemrau in https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/736 -def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps): - x = p.init_latent +@torch.no_grad() +def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps, correction_factor, skip_sdxl_vector): + x = p.init_latent.clone() s_in = x.new_ones([x.shape[0]]) if shared.sd_model.parameterization == "v": @@ -79,43 +73,78 @@ def find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg_scale, steps): for i in trange(1, len(sigmas)): shared.state.sampling_step += 1 - - x_in = torch.cat([x] * 2) sigma_in = torch.cat([sigmas[i - 1] * s_in] * 2) - cond_in = torch.cat([uncond, cond]) - - image_conditioning = torch.cat([p.image_conditioning] * 2) - cond_in = {"c_concat": [image_conditioning], "c_crossattn": [cond_in]} - - c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]] + if shared.sd_model.is_sdxl: + cond_in = {"crossattn": [torch.cat([uncond['crossattn'], cond['crossattn']])], "vector": [torch.cat([uncond['vector'], cond['vector']])]} + else: + cond_in = {"c_concat": [torch.cat([p.image_conditioning] * 2)], "c_crossattn": [torch.cat([uncond, cond])]} if i == 1: t = dnw.sigma_to_t(torch.cat([sigmas[i] * s_in] * 2)) + dt = (sigmas[i] - sigmas[i - 1]) / (2 * sigmas[i]) else: t = dnw.sigma_to_t(sigma_in) + dt = (sigmas[i] - sigmas[i - 1]) / sigmas[i - 1] - eps = shared.sd_model.apply_model(x_in * c_in, t, cond=cond_in) - denoised_uncond, denoised_cond = (x_in + eps * c_out).chunk(2) + noise = noise_from_model(x, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip, skip_sdxl_vector) - denoised = denoised_uncond + (denoised_cond - denoised_uncond) * cfg_scale + if correction_factor > 0: # runs model with previously calculated noise + recalculated_noise = noise_from_model(x + noise, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip, skip_sdxl_vector) + noise = recalculated_noise * correction_factor + noise * (1 - correction_factor) - if i == 1: - d = (x - denoised) / (2 * sigmas[i]) + x += noise + + sd_samplers_common.store_latent(x) + + shared.state.nextjob() + + return x, sigmas[-1] + +@torch.no_grad() +def noise_from_model(x, t, dt, sigma_in, cond_in, cfg_scale, dnw, skip, skip_sdxl_vector): + + if cfg_scale == 1: # Case where denoised_uncond should not be calculated - 50% speedup, also good for sdxl in experiments + x_in = x + sigma_in = sigma_in[1:2] + c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]] + cond_in = {k:[v[0][1:2]] for k, v in cond_in.items()} + if shared.sd_model.is_sdxl: + num_classes_hack = shared.sd_model.model.diffusion_model.num_classes + if skip_sdxl_vector: + shared.sd_model.model.diffusion_model.num_classes = None + cond_in["vector"][0] = None + try: + eps = shared.sd_model.model(x_in * c_in, t[1:2], {"crossattn": cond_in["crossattn"][0], "vector": cond_in["vector"][0]}) + finally: + shared.sd_model.model.diffusion_model.num_classes = num_classes_hack else: - d = (x - denoised) / sigmas[i - 1] + eps = shared.sd_model.apply_model(x_in * c_in, t[1:2], cond=cond_in) - dt = sigmas[i] - sigmas[i - 1] - x = x + d * dt + return -eps * c_out* dt + else : + x_in = torch.cat([x] * 2) - sd_samplers_common.store_latent(x) + c_out, c_in = [K.utils.append_dims(k, x_in.ndim) for k in dnw.get_scalings(sigma_in)[skip:]] - # This shouldn't be necessary, but solved some VRAM issues - del x_in, sigma_in, cond_in, c_out, c_in, t, - del eps, denoised_uncond, denoised_cond, denoised, d, dt + if shared.sd_model.is_sdxl: + num_classes_hack = shared.sd_model.model.diffusion_model.num_classes + if skip_sdxl_vector: + shared.sd_model.model.diffusion_model.num_classes = None + cond_in["vector"][0] = None + try: + eps = shared.sd_model.model(x_in * c_in, t, {"crossattn": cond_in["crossattn"][0], "vector": cond_in["vector"][0]} ) + finally: + shared.sd_model.model.diffusion_model.num_classes = num_classes_hack + else: + eps = shared.sd_model.apply_model(x_in * c_in, t, cond=cond_in) - shared.state.nextjob() + denoised_uncond, denoised_cond = (eps * c_out).chunk(2) + + denoised = denoised_uncond + (denoised_cond - denoised_uncond) * cfg_scale + + return -denoised * dt - return x / sigmas[-1] +Cached = namedtuple("Cached", ["noise", "cfg_scale", "steps", "latent", "original_prompt", "original_negative_prompt", "sigma_adjustment", "second_order_correction", "skip_sdxl_vector"]) class Script(scripts.Script): @@ -133,20 +162,25 @@ def ui(self, is_img2img): * `CFG Scale` should be 2 or lower. ''') - override_sampler = gr.Checkbox(label="Override `Sampling method` to Euler?(this method is built for it)", value=True, elem_id=self.elem_id("override_sampler")) + override_sampler = gr.Checkbox(label="Override `Sampling method` to Euler?(this method is built for it)", value=False, elem_id=self.elem_id("override_sampler")) - override_prompt = gr.Checkbox(label="Override `prompt` to the same value as `original prompt`?(and `negative prompt`)", value=True, elem_id=self.elem_id("override_prompt")) + override_prompt = gr.Checkbox(label="Override `prompt` to the same value as `original prompt`?(and `negative prompt`)", value=False, elem_id=self.elem_id("override_prompt")) original_prompt = gr.Textbox(label="Original prompt", lines=1, elem_id=self.elem_id("original_prompt")) original_negative_prompt = gr.Textbox(label="Original negative prompt", lines=1, elem_id=self.elem_id("original_negative_prompt")) override_steps = gr.Checkbox(label="Override `Sampling Steps` to the same value as `Decode steps`?", value=True, elem_id=self.elem_id("override_steps")) - st = gr.Slider(label="Decode steps", minimum=1, maximum=150, step=1, value=50, elem_id=self.elem_id("st")) + st = gr.Slider(label="Decode steps", minimum=1, maximum=150, step=1, value=20, elem_id=self.elem_id("st")) override_strength = gr.Checkbox(label="Override `Denoising strength` to 1?", value=True, elem_id=self.elem_id("override_strength")) cfg = gr.Slider(label="Decode CFG scale", minimum=0.0, maximum=15.0, step=0.1, value=1.0, elem_id=self.elem_id("cfg")) randomness = gr.Slider(label="Randomness", minimum=0.0, maximum=1.0, step=0.01, value=0.0, elem_id=self.elem_id("randomness")) - sigma_adjustment = gr.Checkbox(label="Sigma adjustment for finding noise for image", value=False, elem_id=self.elem_id("sigma_adjustment")) + sigma_adjustment = gr.Checkbox(label="Sigma adjustment for finding noise for image", value=True, elem_id=self.elem_id("sigma_adjustment")) + second_order_correction = gr.Slider(label="Correct noise by running model again", minimum=0.0, maximum=1.0, step=0.01, value=0.5, elem_id=self.elem_id("second_order_correction"), + info="use 0 (disabled) for original script behaviour, 0.5 reccomended value. Runs the model again to recalculate noise and correct it by given factor. Higher adheres to original image more.") + noise_sigma_intensity = gr.Slider(label="Weight scaling std vs sigma based", minimum=-1.0, maximum=2.0, step=0.01, value=0.5, elem_id=self.elem_id("noise_sigma_intensity"), + info="use 1 for original script behaviour, 0.5 reccomended value. Decides whether to use fixed sigma value or dynamic standard deviation to scale noise. Lower gives softer images.") + skip_sdxl_vector = gr.Checkbox(label="Skip sdxl vectors", info="may cause distortion if false", value=True, elem_id=self.elem_id("skip_sdxl_vector")) return [ info, @@ -154,10 +188,12 @@ def ui(self, is_img2img): override_prompt, original_prompt, original_negative_prompt, override_steps, st, override_strength, - cfg, randomness, sigma_adjustment, + cfg, randomness, sigma_adjustment, second_order_correction, + noise_sigma_intensity, skip_sdxl_vector ] - def run(self, p, _, override_sampler, override_prompt, original_prompt, original_negative_prompt, override_steps, st, override_strength, cfg, randomness, sigma_adjustment): + @torch.no_grad() + def run(self, p, _, override_sampler, override_prompt, original_prompt, original_negative_prompt, override_steps, st, override_strength, cfg, randomness, sigma_adjustment, second_order_correction, noise_sigma_intensity, skip_sdxl_vector): # Override if override_sampler: p.sampler_name = "Euler" @@ -175,33 +211,46 @@ def sample_extra(conditioning, unconditional_conditioning, seeds, subseeds, subs same_params = self.cache is not None and self.cache.cfg_scale == cfg and self.cache.steps == st \ and self.cache.original_prompt == original_prompt \ and self.cache.original_negative_prompt == original_negative_prompt \ - and self.cache.sigma_adjustment == sigma_adjustment + and self.cache.sigma_adjustment == sigma_adjustment \ + and self.cache.second_order_correction == second_order_correction \ + and self.cache.skip_sdxl_vector == skip_sdxl_vector + same_everything = same_params and self.cache.latent.shape == lat.shape and np.abs(self.cache.latent-lat).sum() < 100 + rand_noise = processing.create_random_tensors(p.init_latent.shape[1:], seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, seed_resize_from_h=p.seed_resize_from_h, seed_resize_from_w=p.seed_resize_from_w, p=p) + if same_everything: - rec_noise = self.cache.noise + rec_noise, sigma_val = self.cache.noise else: + # This prevents a crash, because I don't know how to access the underlying .diffusion_model yet when controlnet is enabled. + # modules.sd_unet -> we're good + # scripts.hook -> we're cooked + if "scripts.hook" in str(shared.sd_model.model.diffusion_model.forward.__module__): + print("turn off any controlnets, do 1 pass and then turn controlnet back on to cache noise") + p.steps = 1 + return sd_samplers.create_sampler(p.sampler_name, p.sd_model).sample_img2img(p, p.init_latent, rand_noise, conditioning, unconditional_conditioning, image_conditioning=p.image_conditioning) + shared.state.job_count += 1 cond = p.sd_model.get_learned_conditioning(p.batch_size * [original_prompt]) uncond = p.sd_model.get_learned_conditioning(p.batch_size * [original_negative_prompt]) if sigma_adjustment: - rec_noise = find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg, st) + rec_noise, sigma_val = find_noise_for_image_sigma_adjustment(p, cond, uncond, cfg, st, second_order_correction, skip_sdxl_vector) else: - rec_noise = find_noise_for_image(p, cond, uncond, cfg, st) - self.cache = Cached(rec_noise, cfg, st, lat, original_prompt, original_negative_prompt, sigma_adjustment) + rec_noise, sigma_val = find_noise_for_image(p, cond, uncond, cfg, st, skip_sdxl_vector) + self.cache = Cached((rec_noise, sigma_val), cfg, st, lat, original_prompt, original_negative_prompt, sigma_adjustment, second_order_correction, skip_sdxl_vector) - rand_noise = processing.create_random_tensors(p.init_latent.shape[1:], seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, seed_resize_from_h=p.seed_resize_from_h, seed_resize_from_w=p.seed_resize_from_w, p=p) + rec_noise = rec_noise / (rec_noise.std()*(1 - noise_sigma_intensity) + sigma_val*noise_sigma_intensity) combined_noise = ((1 - randomness) * rec_noise + randomness * rand_noise) / ((randomness**2 + (1-randomness)**2) ** 0.5) sampler = sd_samplers.create_sampler(p.sampler_name, p.sd_model) + p.seed = p.seed + 1 + sigmas = sampler.model_wrap.get_sigmas(p.steps) noise_dt = combined_noise - (p.init_latent / sigmas[0]) - p.seed = p.seed + 1 - return sampler.sample_img2img(p, p.init_latent, noise_dt, conditioning, unconditional_conditioning, image_conditioning=p.image_conditioning) p.sample = sample_extra