stable-diffusion-webui/modules/sd_samplers_cfg_denoiser.py

import torch
from modules import prompt_parser, devices, sd_samplers_common

from modules.shared import opts, state
import modules.shared as shared
from modules.script_callbacks import CFGDenoiserParams, cfg_denoiser_callback
from modules.script_callbacks import CFGDenoisedParams, cfg_denoised_callback
from modules.script_callbacks import AfterCFGCallbackParams, cfg_after_cfg_callback


def catenate_conds(conds):
    if not isinstance(conds[0], dict):
        return torch.cat(conds)

    return {key: torch.cat([x[key] for x in conds]) for key in conds[0].keys()}


def subscript_cond(cond, a, b):
    if not isinstance(cond, dict):
        return cond[a:b]

    return {key: vec[a:b] for key, vec in cond.items()}


def pad_cond(tensor, repeats, empty):
    if not isinstance(tensor, dict):
        return torch.cat([tensor, empty.repeat((tensor.shape[0], repeats, 1))], axis=1)

    tensor['crossattn'] = pad_cond(tensor['crossattn'], repeats, empty)
    return tensor


class CFGDenoiser(torch.nn.Module):
    """
    Classifier free guidance denoiser. A wrapper for stable diffusion model (specifically for unet)
    that can take a noisy picture and produce a noise-free picture using two guidances (prompts)
    instead of one. Originally, the second prompt is just an empty string, but we use non-empty
    negative prompt.
    """

    def __init__(self, sampler):
        super().__init__()
        self.model_wrap = None
        self.mask = None
        self.nmask = None
        self.mask_blend_power = 1
        self.mask_blend_scale = 1
        self.mask_blend_offset = 0
        self.init_latent = None
        self.steps = None
        """number of steps as specified by user in UI"""

        self.total_steps = None
        """expected number of calls to denoiser calculated from self.steps and specifics of the selected sampler"""

        self.step = 0
        self.image_cfg_scale = None
        self.padded_cond_uncond = False
        self.sampler = sampler
        self.model_wrap = None
        self.p = None

        # NOTE: masking before denoising can cause the original latents to be oversmoothed
        # as the original latents do not have noise
        self.mask_before_denoising = False

    @property
    def inner_model(self):
        raise NotImplementedError()

    def combine_denoised(self, x_out, conds_list, uncond, cond_scale):
        denoised_uncond = x_out[-uncond.shape[0]:]
        denoised = torch.clone(denoised_uncond)

        for i, conds in enumerate(conds_list):
            for cond_index, weight in conds:
                denoised[i] += (x_out[cond_index] - denoised_uncond[i]) * (weight * cond_scale)

        return denoised

    def combine_denoised_for_edit_model(self, x_out, cond_scale):
        out_cond, out_img_cond, out_uncond = x_out.chunk(3)
        denoised = out_uncond + cond_scale * (out_cond - out_img_cond) + self.image_cfg_scale * (out_img_cond - out_uncond)

        return denoised

    def get_pred_x0(self, x_in, x_out, sigma):
        return x_out

    def update_inner_model(self):
        self.model_wrap = None

        c, uc = self.p.get_conds()
        self.sampler.sampler_extra_args['cond'] = c
        self.sampler.sampler_extra_args['uncond'] = uc

    def forward(self, x, sigma, uncond, cond, cond_scale, s_min_uncond, image_cond):
        def latent_blend(a, b, t):
            """
            Interpolates two latent image representations according to the parameter t,
            where the interpolated vectors' magnitudes are also interpolated separately.
            The "detail_preservation" factor biases the magnitude interpolation towards
            the larger of the two magnitudes.
            """
            # Record the original latent vector magnitudes.
            # We bring them to a power so that larger magnitudes are favored over smaller ones.
            # 64-bit operations are used here to allow large exponents.
            detail_preservation = 32
            a_magnitude = torch.norm(a, p=2, dim=1).to(torch.float64) ** detail_preservation
            b_magnitude = torch.norm(b, p=2, dim=1).to(torch.float64) ** detail_preservation

            one_minus_t = 1 - t

            # Interpolate the powered magnitudes, then un-power them (bring them back to a power of 1).
            interp_magnitude = (a_magnitude * one_minus_t + b_magnitude * t) ** (1 / detail_preservation)

            # Linearly interpolate the image vectors.
            image_interp = a * one_minus_t + b * t

            # Calculate the magnitude of the interpolated vectors. (We will remove this magnitude.)
            # 64-bit operations are used here to allow large exponents.
            image_interp_magnitude = torch.norm(image_interp, p=2, dim=1).to(torch.float64) + 0.0001

            # Change the linearly interpolated image vectors' magnitudes to the value we want.
            # This is the last 64-bit operation.
            image_interp *= (interp_magnitude / image_interp_magnitude).to(image_interp.dtype)

            return image_interp

        def get_modified_nmask(nmask, _sigma):
            """
            Converts a negative mask representing the transparency of the original latent vectors being overlayed
            to a mask that is scaled according to the denoising strength for this step.

            Where:
                0 = fully opaque, infinite density, fully masked
                1 = fully transparent, zero density, fully unmasked

            We bring this transparency to a power, as this allows one to simulate N number of blending operations
            where N can be any positive real value. Using this one can control the balance of influence between
            the denoiser and the original latents according to the sigma value.

            NOTE: "mask" is not used
            """
            return torch.pow(nmask, (_sigma ** self.mask_blend_power) * self.mask_blend_scale + self.mask_blend_offset)

        if state.interrupted or state.skipped:
            raise sd_samplers_common.InterruptedException

        if sd_samplers_common.apply_refiner(self):
            cond = self.sampler.sampler_extra_args['cond']
            uncond = self.sampler.sampler_extra_args['uncond']

        # at self.image_cfg_scale == 1.0 produced results for edit model are the same as with normal sampling,
        # so is_edit_model is set to False to support AND composition.
        is_edit_model = shared.sd_model.cond_stage_key == "edit" and self.image_cfg_scale is not None and self.image_cfg_scale != 1.0

        conds_list, tensor = prompt_parser.reconstruct_multicond_batch(cond, self.step)
        uncond = prompt_parser.reconstruct_cond_batch(uncond, self.step)

        assert not is_edit_model or all(len(conds) == 1 for conds in conds_list), "AND is not supported for InstructPix2Pix checkpoint (unless using Image CFG scale = 1.0)"

        # Blend in the original latents (before)
        if self.mask_before_denoising and self.mask is not None:
            x = latent_blend(self.init_latent, x, get_modified_nmask(self.nmask, sigma))

        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
            if isinstance(uncond, dict):
                make_condition_dict = lambda c_crossattn, c_concat: {**c_crossattn, "c_concat": [c_concat]}
            else:
                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_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_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)
        x_in = denoiser_params.x
        image_cond_in = denoiser_params.image_cond
        sigma_in = denoiser_params.sigma
        tensor = denoiser_params.text_cond
        uncond = denoiser_params.text_uncond
        skip_uncond = False

        # alternating uncond allows for higher thresholds without the quality loss normally expected from raising it
        if self.step % 2 and s_min_uncond > 0 and sigma[0] < s_min_uncond and not is_edit_model:
            skip_uncond = True
            x_in = x_in[:-batch_size]
            sigma_in = sigma_in[:-batch_size]

        self.padded_cond_uncond = False
        if shared.opts.pad_cond_uncond and tensor.shape[1] != uncond.shape[1]:
            empty = shared.sd_model.cond_stage_model_empty_prompt
            num_repeats = (tensor.shape[1] - uncond.shape[1]) // empty.shape[1]

            if num_repeats < 0:
                tensor = pad_cond(tensor, -num_repeats, empty)
                self.padded_cond_uncond = True
            elif num_repeats > 0:
                uncond = pad_cond(uncond, num_repeats, empty)
                self.padded_cond_uncond = True

        if tensor.shape[1] == uncond.shape[1] or skip_uncond:
            if is_edit_model:
                cond_in = catenate_conds([tensor, uncond, uncond])
            elif skip_uncond:
                cond_in = tensor
            else:
                cond_in = catenate_conds([tensor, uncond])

            if shared.opts.batch_cond_uncond:
                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=make_condition_dict(subscript_cond(cond_in, a, b), image_cond_in[a:b]))
        else:
            x_out = torch.zeros_like(x_in)
            batch_size = batch_size*2 if shared.opts.batch_cond_uncond else batch_size
            for batch_offset in range(0, tensor.shape[0], batch_size):
                a = batch_offset
                b = min(a + batch_size, tensor.shape[0])

                if not is_edit_model:
                    c_crossattn = subscript_cond(tensor, a, b)
                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=make_condition_dict(c_crossattn, image_cond_in[a:b]))

            if not skip_uncond:
                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_image_indexes = [x[0][0] for x in conds_list]
        if skip_uncond:
            fake_uncond = torch.cat([x_out[i:i+1] for i in denoised_image_indexes])
            x_out = torch.cat([x_out, fake_uncond])  # we skipped uncond denoising, so we put cond-denoised image to where the uncond-denoised image should be

        denoised_params = CFGDenoisedParams(x_out, state.sampling_step, state.sampling_steps, self.inner_model)
        cfg_denoised_callback(denoised_params)

        devices.test_for_nans(x_out, "unet")

        if is_edit_model:
            denoised = self.combine_denoised_for_edit_model(x_out, cond_scale)
        elif skip_uncond:
            denoised = self.combine_denoised(x_out, conds_list, uncond, 1.0)
        else:
            denoised = self.combine_denoised(x_out, conds_list, uncond, cond_scale)

        # Blend in the original latents (after)
        if not self.mask_before_denoising and self.mask is not None:
            denoised = latent_blend(self.init_latent, denoised, get_modified_nmask(self.nmask, sigma))

        self.sampler.last_latent = self.get_pred_x0(torch.cat([x_in[i:i + 1] for i in denoised_image_indexes]), torch.cat([x_out[i:i + 1] for i in denoised_image_indexes]), sigma)

        if opts.live_preview_content == "Prompt":
            preview = self.sampler.last_latent
        elif opts.live_preview_content == "Negative prompt":
            preview = self.get_pred_x0(x_in[-uncond.shape[0]:], x_out[-uncond.shape[0]:], sigma)
        else:
            preview = self.get_pred_x0(torch.cat([x_in[i:i+1] for i in denoised_image_indexes]), torch.cat([denoised[i:i+1] for i in denoised_image_indexes]), sigma)

        sd_samplers_common.store_latent(preview)

        after_cfg_callback_params = AfterCFGCallbackParams(denoised, state.sampling_step, state.sampling_steps)
        cfg_after_cfg_callback(after_cfg_callback_params)
        denoised = after_cfg_callback_params.x

        self.step += 1
        return denoised