imaginAIry/imaginairy/api/generate_refiners.py

"""Functions for generating refined images"""

import logging
from contextlib import nullcontext
from typing import Any

from imaginairy.config import CONTROL_CONFIG_SHORTCUTS
from imaginairy.schema import ControlInput, ImaginePrompt, MaskMode
from imaginairy.utils import clear_gpu_cache, seed_everything
from imaginairy.utils.log_utils import ImageLoggingContext

logger = logging.getLogger(__name__)


def generate_single_image(
    prompt: ImaginePrompt,
    debug_img_callback=None,
    progress_img_callback=None,
    progress_img_interval_steps=3,
    progress_img_interval_min_s=0.1,
    add_caption=False,
    return_latent=False,
    dtype=None,
    logging_context: ImageLoggingContext | None = None,
    output_perf=False,
    image_name="",
):
    import torch.nn
    from PIL import Image, ImageOps
    from tqdm import tqdm

    from imaginairy.api.generate import (
        IMAGINAIRY_SAFETY_MODE,
    )
    from imaginairy.enhancers.clip_masking import get_img_mask
    from imaginairy.enhancers.describe_image_blip import generate_caption
    from imaginairy.enhancers.face_restoration_codeformer import enhance_faces
    from imaginairy.enhancers.upscale_realesrgan import upscale_image
    from imaginairy.samplers import SolverName
    from imaginairy.schema import ImagineResult
    from imaginairy.utils import get_device, randn_seeded
    from imaginairy.utils.img_utils import (
        add_caption_to_image,
        combine_image,
        pillow_fit_image_within,
        pillow_img_to_torch_image,
        pillow_mask_to_latent_mask,
    )
    from imaginairy.utils.log_utils import (
        ImageLoggingContext,
        log_img,
        log_latent,
    )
    from imaginairy.utils.model_manager import (
        get_diffusion_model_refiners,
        get_model_default_image_size,
    )
    from imaginairy.utils.outpaint import (
        outpaint_arg_str_parse,
        prepare_image_for_outpaint,
    )
    from imaginairy.utils.safety import create_safety_score
    from imaginairy.vendored.refiners.foundationals.latent_diffusion.schedulers import (
        DDIM,
        DPMSolver,
    )

    if dtype is None:
        dtype = torch.float16

    get_device()
    clear_gpu_cache()
    prompt = prompt.make_concrete_copy()

    if not logging_context:

        def latent_logger(latents):
            progress_latents.append(latents)

        lc = ImageLoggingContext(
            prompt=prompt,
            debug_img_callback=debug_img_callback,
            progress_img_callback=progress_img_callback,
            progress_img_interval_steps=progress_img_interval_steps,
            progress_img_interval_min_s=progress_img_interval_min_s,
            progress_latent_callback=latent_logger
            if prompt.collect_progress_latents
            else None,
        )
        _context: Any = lc
    else:
        lc = logging_context
        _context = nullcontext()
    with _context:
        with lc.timing("model-load"):
            sd = get_diffusion_model_refiners(
                weights_config=prompt.model_weights,
                for_inpainting=prompt.should_use_inpainting
                and prompt.inpaint_method == "finetune",
                dtype=dtype,
            )
        lc.model = sd
        seed_everything(prompt.seed)
        downsampling_factor = 8
        latent_channels = 4
        batch_size = 1

        mask_image = None
        mask_image_orig = None

        sd.set_tile_mode(prompt.tile_mode)

        result_images: dict[str, torch.Tensor | None | Image.Image] = {}
        progress_latents: list[torch.Tensor] = []
        first_step = 0
        mask_grayscale = None

        shape = [
            batch_size,
            latent_channels,
            prompt.height // downsampling_factor,
            prompt.width // downsampling_factor,
        ]

        init_latent = None
        noise_step = None

        control_modes = []
        control_inputs = prompt.control_inputs or []
        control_inputs = control_inputs.copy()

        if control_inputs:
            control_modes = [c.mode for c in prompt.control_inputs]

        if prompt.init_image:
            starting_image = prompt.init_image
            assert prompt.init_image_strength is not None
            first_step = int(prompt.steps * prompt.init_image_strength)
            # noise_step = int((prompt.steps - 1) * prompt.init_image_strength)

            if prompt.mask_prompt:
                mask_image, mask_grayscale = get_img_mask(
                    starting_image, prompt.mask_prompt, threshold=0.1
                )
            elif prompt.mask_image:
                mask_image = prompt.mask_image.convert("L")

            if prompt.outpaint:
                outpaint_kwargs = outpaint_arg_str_parse(prompt.outpaint)
                starting_image, mask_image = prepare_image_for_outpaint(
                    starting_image, mask_image, **outpaint_kwargs
                )

            init_image = pillow_fit_image_within(
                starting_image,
                max_height=prompt.height,
                max_width=prompt.width,
            )
            init_image_t = pillow_img_to_torch_image(init_image)
            init_image_t = init_image_t.to(device=sd.lda.device, dtype=sd.lda.dtype)
            init_latent = sd.lda.encode(init_image_t)

            shape = list(init_latent.shape)

            log_latent(init_latent, "init_latent")

            if mask_image is not None:
                mask_image = pillow_fit_image_within(
                    mask_image,
                    max_height=prompt.height,
                    max_width=prompt.width,
                    convert="L",
                )

                log_img(mask_image, "init mask")

                if prompt.mask_mode == MaskMode.REPLACE:
                    mask_image = ImageOps.invert(mask_image)

                mask_image_orig = mask_image
                log_img(mask_image, "latent_mask")
                pillow_mask_to_latent_mask(
                    mask_image, downsampling_factor=downsampling_factor
                ).to(get_device())
                if prompt.inpaint_method == "controlnet":
                    result_images["control-inpaint"] = mask_image
                    control_inputs.append(
                        ControlInput(mode="inpaint", image=mask_image)
                    )

        seed_everything(prompt.seed)
        assert prompt.seed is not None

        noise = randn_seeded(seed=prompt.seed, size=shape).to(
            sd.unet.device, dtype=sd.unet.dtype
        )
        noised_latent = noise
        controlnets = []

        if control_modes:
            with lc.timing("control-image-prep"):
                for control_input in control_inputs:
                    (
                        controlnet,
                        control_image_t,
                        control_image_disp,
                    ) = prep_control_input(
                        control_input=control_input,
                        sd=sd,
                        init_image_t=init_image_t,
                        fit_width=prompt.width,
                        fit_height=prompt.height,
                    )
                    result_images[f"control-{control_input.mode}"] = control_image_disp
                    controlnets.append((controlnet, control_image_t))

        if prompt.allow_compose_phase:
            with lc.timing("composition"):
                cutoff_size = get_model_default_image_size(prompt.model_architecture)
                cutoff_size = (int(cutoff_size[0] * 1.00), int(cutoff_size[1] * 1.00))
                compose_kwargs = {
                    "prompt": prompt,
                    "target_height": prompt.height,
                    "target_width": prompt.width,
                    "cutoff": cutoff_size,
                    "dtype": dtype,
                }

                if prompt.init_image:
                    compose_kwargs.update(
                        {
                            "target_height": init_image.height,
                            "target_width": init_image.width,
                        }
                    )
                comp_image, comp_img_orig = _generate_composition_image(
                    **compose_kwargs, logging_context=lc
                )

                if comp_image is not None:
                    prompt.fix_faces = False  # done in composition
                    result_images["composition"] = comp_img_orig
                    result_images["composition-upscaled"] = comp_image
                    composition_strength = prompt.composition_strength
                    first_step = int((prompt.steps) * composition_strength)
                    noise_step = int((prompt.steps - 1) * composition_strength)
                    log_img(comp_img_orig, "comp_image")
                    log_img(comp_image, "comp_image_upscaled")
                    comp_image_t = pillow_img_to_torch_image(comp_image)
                    comp_image_t = comp_image_t.to(sd.lda.device, dtype=sd.lda.dtype)
                    init_latent = sd.lda.encode(comp_image_t)
                    compose_control_inputs: list[ControlInput]
                    if prompt.model_weights.architecture.primary_alias in (
                        "sdxl",
                        "sdxlinpaint",
                    ):
                        compose_control_inputs = []
                    else:
                        compose_control_inputs = [
                            ControlInput(mode="details", image=comp_image, strength=1),
                        ]
                    for control_input in compose_control_inputs:
                        (
                            controlnet,
                            control_image_t,
                            control_image_disp,
                        ) = prep_control_input(
                            control_input=control_input,
                            sd=sd,
                            init_image_t=None,
                            fit_width=prompt.width,
                            fit_height=prompt.height,
                        )
                        result_images[f"control-{control_input.mode}"] = (
                            control_image_disp
                        )
                        controlnets.append((controlnet, control_image_t))

        for controlnet, control_image_t in controlnets:
            controlnet.set_controlnet_condition(
                control_image_t.to(device=sd.unet.device, dtype=sd.unet.dtype)
            )
            controlnet.inject()
        if prompt.solver_type.lower() == SolverName.DPMPP:
            sd.scheduler = DPMSolver(num_inference_steps=prompt.steps)
        elif prompt.solver_type.lower() == SolverName.DDIM:
            sd.scheduler = DDIM(num_inference_steps=prompt.steps)
        else:
            msg = f"Unknown solver type: {prompt.solver_type}"
            raise ValueError(msg)
        sd.scheduler.to(device=sd.unet.device, dtype=sd.unet.dtype)
        sd.set_inference_steps(prompt.steps, first_step=first_step)

        if hasattr(sd, "mask_latents") and mask_image is not None:
            sd.set_inpainting_conditions(
                target_image=init_image,
                mask=ImageOps.invert(mask_image),
                latents_size=shape[-2:],
            )
            sd.target_image_latents = sd.target_image_latents.to(
                dtype=sd.unet.dtype, device=sd.unet.device
            )
            sd.mask_latents = sd.mask_latents.to(
                dtype=sd.unet.dtype, device=sd.unet.device
            )

        if init_latent is not None:
            noise_step = noise_step if noise_step is not None else first_step
            if first_step >= len(sd.scheduler.all_steps):
                noised_latent = init_latent
            else:
                noised_latent = sd.scheduler.add_noise(
                    x=init_latent, noise=noise, step=sd.scheduler.all_steps[noise_step]
                )

        with lc.timing("text-conditioning"):
            text_conditioning_kwargs = sd.calculate_text_conditioning_kwargs(
                positive_prompts=prompt.prompts,
                negative_prompts=prompt.negative_prompt,
                positive_conditioning_override=prompt.conditioning,
            )

            for k, v in text_conditioning_kwargs.items():
                text_conditioning_kwargs[k] = v.to(
                    device=sd.unet.device, dtype=sd.unet.dtype
                )
        x = noised_latent
        x = x.to(device=sd.unet.device, dtype=sd.unet.dtype)
        clear_gpu_cache()

        with lc.timing("unet"):
            for step in tqdm(
                sd.steps, bar_format="    {l_bar}{bar}{r_bar}", leave=False
            ):
                log_latent(x, "noisy_latent")
                x = sd(
                    x,
                    step=step,
                    condition_scale=prompt.prompt_strength,
                    **text_conditioning_kwargs,
                )
                if lc.progress_latent_callback:
                    lc.progress_latent_callback(x)
            # trying to clear memory. not sure if this helps
            sd.unet.set_context(context="self_attention_map", value={})
            sd.unet._reset_context()
            clear_gpu_cache()

        logger.debug("Decoding image")
        if x.device != sd.lda.device:
            sd.lda.to(x.device)
        clear_gpu_cache()
        with lc.timing("decode-img"):
            gen_img = sd.lda.decode_latents(x.to(dtype=sd.lda.dtype))

        if mask_image_orig and init_image:
            with lc.timing("combine-image"):
                result_images["pre-reconstitution"] = gen_img
                mask_final = mask_image_orig.copy()
                # mask_final = ImageOps.invert(mask_final)

                log_img(mask_final, "reconstituting mask")
                # gen_img = Image.composite(gen_img, init_image, mask_final)
                gen_img = combine_image(
                    original_img=init_image,
                    generated_img=gen_img,
                    mask_img=mask_final,
                )
                log_img(gen_img, "reconstituted image")

        upscaled_img = None
        rebuilt_orig_img = None

        if add_caption:
            with lc.timing("caption-img"):
                caption = generate_caption(gen_img)
                logger.info(f"Generated caption: {caption}")

        with lc.timing("safety-filter"):
            safety_score = create_safety_score(
                gen_img,
                safety_mode=IMAGINAIRY_SAFETY_MODE,
            )
        if safety_score.is_filtered:
            progress_latents.clear()
        if not safety_score.is_filtered:
            if prompt.fix_faces:
                with lc.timing("face-enhancement"):
                    logger.info("Fixing 😊 's in 🖼  using CodeFormer...")
                    with lc.timing("face-enhancement"):
                        gen_img = enhance_faces(
                            gen_img, fidelity=prompt.fix_faces_fidelity
                        )
            if prompt.upscale:
                with lc.timing("upscaling"):
                    logger.info("Upscaling 🖼  using real-ESRGAN...")
                    with lc.timing("upscaling"):
                        upscaled_img = upscale_image(gen_img)

            # put the newly generated patch back into the original, full-size image
            if prompt.mask_modify_original and mask_image_orig and starting_image:
                logger.info("Combining inpainting with original image...")
                img_to_add_back_to_original = upscaled_img if upscaled_img else gen_img
                rebuilt_orig_img = combine_image(
                    original_img=starting_image,
                    generated_img=img_to_add_back_to_original,
                    mask_img=mask_image_orig,
                )

            if prompt.caption_text:
                caption_text = prompt.caption_text.format(prompt=prompt.prompt_text)
                add_caption_to_image(gen_img, caption_text)

        # todo: do something smarter
        result_images.update(
            {
                "upscaled": upscaled_img,
                "modified_original": rebuilt_orig_img,
                "mask_binary": mask_image_orig,
                "mask_grayscale": mask_grayscale,
            }
        )

        result = ImagineResult(
            img=gen_img,
            prompt=prompt,
            is_nsfw=safety_score.is_nsfw,
            safety_score=safety_score,
            result_images=result_images,
            performance_stats=lc.get_performance_stats(),
            progress_latents=progress_latents,
        )

        _most_recent_result = result
        _image_name = f"{image_name} " if image_name else ""
        logger.info(f"Generated {_image_name}image in {result.total_time():.1f}s")

        if result.performance_stats:
            log = logger.info if output_perf else logger.debug
            log(f"   Timings: {result.timings_str()}")
            if torch.cuda.is_available():
                log(f"   Peak VRAM: {result.gpu_str('memory_peak')}")
                log(f"   Peak VRAM Delta: {result.gpu_str('memory_peak_delta')}")
                log(f"   Ending VRAM: {result.gpu_str('memory_end')}")
        for controlnet, _ in controlnets:
            controlnet.eject()
        clear_gpu_cache()
        return result


def prep_control_input(
    control_input: ControlInput, sd, init_image_t, fit_width, fit_height
):
    from PIL import ImageOps

    from imaginairy.utils import get_device
    from imaginairy.utils.img_utils import (
        pillow_fit_image_within,
        pillow_img_to_torch_image,
    )
    from imaginairy.utils.log_utils import (
        log_img,
    )

    if control_input.image_raw is not None:
        control_image = control_input.image_raw
    elif control_input.image is not None:
        control_image = control_input.image
    else:
        raise ValueError("No control image found")
    assert control_image is not None
    control_image = control_image.convert("RGB")
    log_img(control_image, "control_image_input")
    control_image_input = pillow_fit_image_within(
        control_image,
        max_height=fit_height,
        max_width=fit_width,
    )
    if control_input.mode == "inpaint":
        control_image_input = ImageOps.invert(control_image_input)

    control_image_input_t = pillow_img_to_torch_image(control_image_input)
    control_image_input_t = control_image_input_t.to(get_device())

    if control_input.image_raw is None:
        from imaginairy.img_processors.control_modes import CONTROL_MODES

        control_prep_function = CONTROL_MODES[control_input.mode]
        if control_input.mode == "inpaint":
            control_image_t = control_prep_function(  # type: ignore
                control_image_input_t, init_image_t
            )
        else:
            control_image_t = control_prep_function(control_image_input_t)  # type: ignore
    else:
        control_image_t = (control_image_input_t + 1) / 2

    control_image_disp = control_image_t * 2 - 1

    log_img(control_image_disp, "control_image")

    if len(control_image_t.shape) == 3:
        raise ValueError("Control image must be 4D")

    if control_image_t.shape[1] != 3:
        raise ValueError("Control image must have 3 channels")

    if (
        control_input.mode != "inpaint"
        and control_image_t.min() < 0
        or control_image_t.max() > 1
    ):
        msg = f"Control image must be in [0, 1] but we received {control_image_t.min()} and {control_image_t.max()}"
        raise ValueError(msg)

    if control_image_t.max() == control_image_t.min():
        msg = f"No control signal found in control image {control_input.mode}."
        raise ValueError(msg)

    control_config = CONTROL_CONFIG_SHORTCUTS.get(control_input.mode, None)
    if not control_config:
        msg = f"Unknown control mode: {control_input.mode}"
        raise ValueError(msg)
    from imaginairy.vendored.refiners.foundationals.latent_diffusion import (
        SD1ControlnetAdapter,
    )

    controlnet = SD1ControlnetAdapter(  # type: ignore
        name=control_input.mode,
        target=sd.unet,
        weights_location=control_config.weights_location,
    )

    controlnet.set_scale(control_input.strength)
    control_image_t = control_image_t.to(device=sd.unet.device, dtype=sd.unet.dtype)
    return controlnet, control_image_t, control_image_disp


def _generate_composition_image(
    prompt,
    target_height,
    target_width,
    cutoff: tuple[int, int] = (512, 512),
    dtype=None,
    logging_context=None,
):
    from PIL import Image

    from imaginairy.utils import default, get_default_dtype
    from imaginairy.utils.img_utils import calc_scale_to_fit_within
    from imaginairy.utils.named_resolutions import normalize_image_size

    cutoff = normalize_image_size(cutoff)
    if prompt.width <= cutoff[0] and prompt.height <= cutoff[1]:
        return None, None

    dtype = default(dtype, get_default_dtype)

    shrink_scale = calc_scale_to_fit_within(
        height=prompt.height,
        width=prompt.width,
        max_size=cutoff,
    )

    composition_prompt = prompt.full_copy(
        deep=True,
        update={
            "size": (
                int(prompt.width * shrink_scale),
                int(prompt.height * shrink_scale),
            ),
            "steps": None,
            "upscale": False,
            "fix_faces": False,
            "mask_modify_original": False,
            "allow_compose_phase": False,
            "caption_text": None,
        },
    )

    result = generate_single_image(
        composition_prompt,
        dtype=dtype,
        logging_context=logging_context,
        output_perf=False,
        image_name="composition",
    )
    img = result.images["generated"]
    while img.width < target_width:
        from imaginairy.enhancers.upscale_realesrgan import upscale_image

        if prompt.fix_faces:
            from imaginairy.enhancers.face_restoration_codeformer import enhance_faces

            with logging_context.timing("face-enhancement"):
                logger.info("Fixing 😊 's in 🖼  using CodeFormer...")
                img = enhance_faces(img, fidelity=prompt.fix_faces_fidelity)
        with logging_context.timing("upscaling"):
            img = upscale_image(img, ultrasharp=True)

    img = img.resize(
        (target_width, target_height),
        resample=Image.Resampling.LANCZOS,
    )

    return img, result.images["generated"]