mirror of
https://github.com/invoke-ai/InvokeAI.git
synced 2026-04-23 03:00:31 -04:00
tidy(ui): remove unused stuff
This commit is contained in:
psychedelicious
@@ -1,166 +0,0 @@
|
||||
import type { RootState } from 'app/store/store';
|
||||
import { deepClone } from 'common/util/deepClone';
|
||||
import { roundToMultiple } from 'common/util/roundDownToMultiple';
|
||||
import { selectOptimalDimension } from 'features/controlLayers/store/selectors';
|
||||
import {
|
||||
DENOISE_LATENTS_HRF,
|
||||
ESRGAN_HRF,
|
||||
IMAGE_TO_LATENTS_HRF,
|
||||
LATENTS_TO_IMAGE_HRF_HR,
|
||||
LATENTS_TO_IMAGE_HRF_LR,
|
||||
NOISE_HRF,
|
||||
RESIZE_HRF,
|
||||
} from 'features/nodes/util/graph/constants';
|
||||
import type { Graph } from 'features/nodes/util/graph/generation/Graph';
|
||||
import { getBoardField } from 'features/nodes/util/graph/graphBuilderUtils';
|
||||
import type { Invocation } from 'services/api/types';
|
||||
|
||||
/**
|
||||
* Calculates the new resolution for high-resolution features (HRF) based on base model type.
|
||||
* Adjusts the width and height to maintain the aspect ratio and constrains them by the model's dimension limits,
|
||||
* rounding down to the nearest multiple of 8.
|
||||
*
|
||||
* @param {number} optimalDimension The optimal dimension for the base model.
|
||||
* @param {number} width The current width to be adjusted for HRF.
|
||||
* @param {number} height The current height to be adjusted for HRF.
|
||||
* @return {{newWidth: number, newHeight: number}} The new width and height, adjusted and rounded as needed.
|
||||
*/
|
||||
function calculateHrfRes(
|
||||
optimalDimension: number,
|
||||
width: number,
|
||||
height: number
|
||||
): { newWidth: number; newHeight: number } {
|
||||
const aspect = width / height;
|
||||
|
||||
const minDimension = Math.floor(optimalDimension * 0.5);
|
||||
const modelArea = optimalDimension * optimalDimension; // Assuming square images for model_area
|
||||
|
||||
let initWidth;
|
||||
let initHeight;
|
||||
|
||||
if (aspect > 1.0) {
|
||||
initHeight = Math.max(minDimension, Math.sqrt(modelArea / aspect));
|
||||
initWidth = initHeight * aspect;
|
||||
} else {
|
||||
initWidth = Math.max(minDimension, Math.sqrt(modelArea * aspect));
|
||||
initHeight = initWidth / aspect;
|
||||
}
|
||||
// Cap initial height and width to final height and width.
|
||||
initWidth = Math.min(width, initWidth);
|
||||
initHeight = Math.min(height, initHeight);
|
||||
|
||||
const newWidth = roundToMultiple(Math.floor(initWidth), 8);
|
||||
const newHeight = roundToMultiple(Math.floor(initHeight), 8);
|
||||
|
||||
return { newWidth, newHeight };
|
||||
}
|
||||
|
||||
/**
|
||||
* Adds HRF to the graph.
|
||||
* @param state The root redux state
|
||||
* @param g The graph to add HRF to
|
||||
* @param denoise The denoise node
|
||||
* @param noise The noise node
|
||||
* @param l2i The l2i node
|
||||
* @param vaeSource The VAE source node (may be a model loader, VAE loader, or seamless node)
|
||||
* @returns The HRF image output node.
|
||||
*/
|
||||
export const addHRF = (
|
||||
state: RootState,
|
||||
g: Graph,
|
||||
denoise: Invocation<'denoise_latents'>,
|
||||
noise: Invocation<'noise'>,
|
||||
l2i: Invocation<'l2i'>,
|
||||
vaeSource: Invocation<'vae_loader'> | Invocation<'main_model_loader'> | Invocation<'seamless'>
|
||||
): Invocation<'l2i'> => {
|
||||
const { hrfStrength, hrfEnabled, hrfMethod } = state.hrf;
|
||||
const { width, height } = state.canvasV2.document;
|
||||
const optimalDimension = selectOptimalDimension(state);
|
||||
const { newWidth: hrfWidth, newHeight: hrfHeight } = calculateHrfRes(optimalDimension, width, height);
|
||||
|
||||
// Change height and width of original noise node to initial resolution.
|
||||
if (noise) {
|
||||
noise.width = hrfWidth;
|
||||
noise.height = hrfHeight;
|
||||
}
|
||||
|
||||
// Define new nodes and their connections, roughly in order of operations.
|
||||
const l2iHrfLR = g.addNode({ type: 'l2i', id: LATENTS_TO_IMAGE_HRF_LR, fp32: l2i.fp32 });
|
||||
g.addEdge(denoise, 'latents', l2iHrfLR, 'latents');
|
||||
g.addEdge(vaeSource, 'vae', l2iHrfLR, 'vae');
|
||||
|
||||
const resizeHrf = g.addNode({
|
||||
id: RESIZE_HRF,
|
||||
type: 'img_resize',
|
||||
width: width,
|
||||
height: height,
|
||||
});
|
||||
|
||||
if (hrfMethod === 'ESRGAN') {
|
||||
let model_name: Invocation<'esrgan'>['model_name'] = 'RealESRGAN_x2plus.pth';
|
||||
if ((width * height) / (hrfWidth * hrfHeight) > 2) {
|
||||
model_name = 'RealESRGAN_x4plus.pth';
|
||||
}
|
||||
const esrganHrf = g.addNode({ id: ESRGAN_HRF, type: 'esrgan', model_name });
|
||||
g.addEdge(l2iHrfLR, 'image', esrganHrf, 'image');
|
||||
g.addEdge(esrganHrf, 'image', resizeHrf, 'image');
|
||||
} else {
|
||||
g.addEdge(l2iHrfLR, 'image', resizeHrf, 'image');
|
||||
}
|
||||
|
||||
const noiseHrf = g.addNode({
|
||||
type: 'noise',
|
||||
id: NOISE_HRF,
|
||||
seed: noise.seed,
|
||||
use_cpu: noise.use_cpu,
|
||||
});
|
||||
g.addEdge(resizeHrf, 'height', noiseHrf, 'height');
|
||||
g.addEdge(resizeHrf, 'width', noiseHrf, 'width');
|
||||
|
||||
const i2lHrf = g.addNode({ type: 'i2l', id: IMAGE_TO_LATENTS_HRF });
|
||||
g.addEdge(vaeSource, 'vae', i2lHrf, 'vae');
|
||||
g.addEdge(resizeHrf, 'image', i2lHrf, 'image');
|
||||
|
||||
const denoiseHrf = g.addNode({
|
||||
type: 'denoise_latents',
|
||||
id: DENOISE_LATENTS_HRF,
|
||||
cfg_scale: denoise.cfg_scale,
|
||||
scheduler: denoise.scheduler,
|
||||
steps: denoise.steps,
|
||||
denoising_start: 1 - hrfStrength,
|
||||
denoising_end: 1,
|
||||
});
|
||||
g.addEdge(i2lHrf, 'latents', denoiseHrf, 'latents');
|
||||
g.addEdge(noiseHrf, 'noise', denoiseHrf, 'noise');
|
||||
|
||||
// Copy edges to the original denoise into the new denoise
|
||||
g.getEdgesTo(denoise, ['control', 'ip_adapter', 'unet', 'positive_conditioning', 'negative_conditioning']).forEach(
|
||||
(edge) => {
|
||||
const clone = deepClone(edge);
|
||||
clone.destination.node_id = denoiseHrf.id;
|
||||
g.addEdgeFromObj(clone);
|
||||
}
|
||||
);
|
||||
|
||||
// The original l2i node is unnecessary now, remove it
|
||||
g.deleteNode(l2i.id);
|
||||
|
||||
const l2iHrfHR = g.addNode({
|
||||
type: 'l2i',
|
||||
id: LATENTS_TO_IMAGE_HRF_HR,
|
||||
fp32: l2i.fp32,
|
||||
is_intermediate: false,
|
||||
board: getBoardField(state),
|
||||
});
|
||||
g.addEdge(vaeSource, 'vae', l2iHrfHR, 'vae');
|
||||
g.addEdge(denoiseHrf, 'latents', l2iHrfHR, 'latents');
|
||||
|
||||
g.upsertMetadata({
|
||||
hrf_strength: hrfStrength,
|
||||
hrf_enabled: hrfEnabled,
|
||||
hrf_method: hrfMethod,
|
||||
});
|
||||
g.setMetadataReceivingNode(l2iHrfHR);
|
||||
|
||||
return l2iHrfHR;
|
||||
};
|
||||
@@ -1,5 +0,0 @@
|
||||
import type { CanvasRasterLayerState } from 'features/controlLayers/store/types';
|
||||
|
||||
export const isValidLayer = (layer: CanvasRasterLayerState) => {
|
||||
return layer.isEnabled && layer.objects.length > 0;
|
||||
};
|
||||
@@ -1,197 +0,0 @@
|
||||
import type { RootState } from 'app/store/store';
|
||||
import type { CanvasManager } from 'features/controlLayers/konva/CanvasManager';
|
||||
import { fetchModelConfigWithTypeGuard } from 'features/metadata/util/modelFetchingHelpers';
|
||||
import {
|
||||
LATENTS_TO_IMAGE,
|
||||
NEGATIVE_CONDITIONING,
|
||||
NEGATIVE_CONDITIONING_COLLECT,
|
||||
NOISE,
|
||||
POSITIVE_CONDITIONING,
|
||||
POSITIVE_CONDITIONING_COLLECT,
|
||||
SDXL_CONTROL_LAYERS_GRAPH,
|
||||
SDXL_DENOISE_LATENTS,
|
||||
SDXL_MODEL_LOADER,
|
||||
VAE_LOADER,
|
||||
} from 'features/nodes/util/graph/constants';
|
||||
import { addControlAdapters } from 'features/nodes/util/graph/generation/addControlAdapters';
|
||||
import { addIPAdapters } from 'features/nodes/util/graph/generation/addIPAdapters';
|
||||
import { addNSFWChecker } from 'features/nodes/util/graph/generation/addNSFWChecker';
|
||||
import { addSDXLLoRAs } from 'features/nodes/util/graph/generation/addSDXLLoRAs';
|
||||
import { addSDXLRefiner } from 'features/nodes/util/graph/generation/addSDXLRefiner';
|
||||
import { addSeamless } from 'features/nodes/util/graph/generation/addSeamless';
|
||||
import { addWatermarker } from 'features/nodes/util/graph/generation/addWatermarker';
|
||||
import { Graph } from 'features/nodes/util/graph/generation/Graph';
|
||||
import { getBoardField, getPresetModifiedPrompts , getSizes } from 'features/nodes/util/graph/graphBuilderUtils';
|
||||
import type { Invocation, NonNullableGraph } from 'services/api/types';
|
||||
import { isNonRefinerMainModelConfig } from 'services/api/types';
|
||||
import { assert } from 'tsafe';
|
||||
|
||||
import { addRegions } from './addRegions';
|
||||
|
||||
export const buildImageToImageSDXLGraph = async (
|
||||
state: RootState,
|
||||
manager: CanvasManager
|
||||
): Promise<NonNullableGraph> => {
|
||||
const { bbox, params } = state.canvasV2;
|
||||
const {
|
||||
model,
|
||||
cfgScale: cfg_scale,
|
||||
cfgRescaleMultiplier: cfg_rescale_multiplier,
|
||||
scheduler,
|
||||
seed,
|
||||
steps,
|
||||
shouldUseCpuNoise,
|
||||
vaePrecision,
|
||||
vae,
|
||||
refinerModel,
|
||||
refinerStart,
|
||||
img2imgStrength,
|
||||
} = params;
|
||||
|
||||
assert(model, 'No model found in state');
|
||||
|
||||
const { positivePrompt, negativePrompt, positiveStylePrompt, negativeStylePrompt } = getPresetModifiedPrompts(state);
|
||||
const { originalSize, scaledSize } = getSizes(bbox);
|
||||
|
||||
|
||||
|
||||
const g = new Graph(SDXL_CONTROL_LAYERS_GRAPH);
|
||||
const modelLoader = g.addNode({
|
||||
type: 'sdxl_model_loader',
|
||||
id: SDXL_MODEL_LOADER,
|
||||
model,
|
||||
});
|
||||
const posCond = g.addNode({
|
||||
type: 'sdxl_compel_prompt',
|
||||
id: POSITIVE_CONDITIONING,
|
||||
prompt: positivePrompt,
|
||||
style: positiveStylePrompt,
|
||||
});
|
||||
const posCondCollect = g.addNode({
|
||||
type: 'collect',
|
||||
id: POSITIVE_CONDITIONING_COLLECT,
|
||||
});
|
||||
const negCond = g.addNode({
|
||||
type: 'sdxl_compel_prompt',
|
||||
id: NEGATIVE_CONDITIONING,
|
||||
prompt: negativePrompt,
|
||||
style: negativeStylePrompt,
|
||||
});
|
||||
const negCondCollect = g.addNode({
|
||||
type: 'collect',
|
||||
id: NEGATIVE_CONDITIONING_COLLECT,
|
||||
});
|
||||
const noise = g.addNode({
|
||||
type: 'noise',
|
||||
id: NOISE,
|
||||
seed,
|
||||
width: scaledSize.width,
|
||||
height: scaledSize.height,
|
||||
use_cpu: shouldUseCpuNoise,
|
||||
});
|
||||
const denoise = g.addNode({
|
||||
type: 'denoise_latents',
|
||||
id: SDXL_DENOISE_LATENTS,
|
||||
cfg_scale,
|
||||
cfg_rescale_multiplier,
|
||||
scheduler,
|
||||
steps,
|
||||
denoising_start: refinerModel ? Math.min(refinerStart, 1 - img2imgStrength) : 1 - img2imgStrength,
|
||||
denoising_end: refinerModel ? refinerStart : 1,
|
||||
});
|
||||
const l2i = g.addNode({
|
||||
type: 'l2i',
|
||||
id: LATENTS_TO_IMAGE,
|
||||
fp32: vaePrecision === 'fp32',
|
||||
board: getBoardField(state),
|
||||
// This is the terminal node and must always save to gallery.
|
||||
is_intermediate: false,
|
||||
use_cache: false,
|
||||
});
|
||||
const vaeLoader =
|
||||
vae?.base === model.base
|
||||
? g.addNode({
|
||||
type: 'vae_loader',
|
||||
id: VAE_LOADER,
|
||||
vae_model: vae,
|
||||
})
|
||||
: null;
|
||||
|
||||
let imageOutput: Invocation<'l2i'> | Invocation<'img_nsfw'> | Invocation<'img_watermark'> | Invocation<'img_resize'> =
|
||||
l2i;
|
||||
|
||||
g.addEdge(modelLoader, 'unet', denoise, 'unet');
|
||||
g.addEdge(modelLoader, 'clip', posCond, 'clip');
|
||||
g.addEdge(modelLoader, 'clip', negCond, 'clip');
|
||||
g.addEdge(modelLoader, 'clip2', posCond, 'clip2');
|
||||
g.addEdge(modelLoader, 'clip2', negCond, 'clip2');
|
||||
g.addEdge(posCond, 'conditioning', posCondCollect, 'item');
|
||||
g.addEdge(negCond, 'conditioning', negCondCollect, 'item');
|
||||
g.addEdge(posCondCollect, 'collection', denoise, 'positive_conditioning');
|
||||
g.addEdge(negCondCollect, 'collection', denoise, 'negative_conditioning');
|
||||
g.addEdge(noise, 'noise', denoise, 'noise');
|
||||
g.addEdge(denoise, 'latents', l2i, 'latents');
|
||||
|
||||
const modelConfig = await fetchModelConfigWithTypeGuard(model.key, isNonRefinerMainModelConfig);
|
||||
assert(modelConfig.base === 'sdxl');
|
||||
|
||||
g.upsertMetadata({
|
||||
generation_mode: 'sdxl_txt2img',
|
||||
cfg_scale,
|
||||
cfg_rescale_multiplier,
|
||||
width: scaledSize.width,
|
||||
height: scaledSize.height,
|
||||
positive_prompt: positivePrompt,
|
||||
negative_prompt: negativePrompt,
|
||||
model: Graph.getModelMetadataField(modelConfig),
|
||||
seed,
|
||||
steps,
|
||||
rand_device: shouldUseCpuNoise ? 'cpu' : 'cuda',
|
||||
scheduler,
|
||||
positive_style_prompt: positiveStylePrompt,
|
||||
negative_style_prompt: negativeStylePrompt,
|
||||
vae: vae ?? undefined,
|
||||
});
|
||||
|
||||
const seamless = addSeamless(state, g, denoise, modelLoader, vaeLoader);
|
||||
|
||||
addSDXLLoRAs(state, g, denoise, modelLoader, seamless, posCond, negCond);
|
||||
|
||||
// We might get the VAE from the main model, custom VAE, or seamless node.
|
||||
const vaeSource = seamless ?? vaeLoader ?? modelLoader;
|
||||
g.addEdge(vaeSource, 'vae', l2i, 'vae');
|
||||
|
||||
// Add Refiner if enabled
|
||||
if (refinerModel) {
|
||||
await addSDXLRefiner(state, g, denoise, seamless, posCond, negCond, l2i);
|
||||
}
|
||||
|
||||
|
||||
|
||||
const _addedCAs = addControlAdapters(state.canvasV2.controlAdapters.entities, g, denoise, modelConfig.base);
|
||||
const _addedIPAs = addIPAdapters(state.canvasV2.ipAdapters.entities, g, denoise, modelConfig.base);
|
||||
const _addedRegions = await addRegions(
|
||||
manager,
|
||||
state.canvasV2.regions.entities,
|
||||
g,
|
||||
state.canvasV2.document,
|
||||
state.canvasV2.bbox,
|
||||
modelConfig.base,
|
||||
denoise,
|
||||
posCond,
|
||||
negCond,
|
||||
posCondCollect,
|
||||
negCondCollect
|
||||
);
|
||||
|
||||
if (state.system.shouldUseNSFWChecker) {
|
||||
imageOutput = addNSFWChecker(g, imageOutput);
|
||||
}
|
||||
|
||||
if (state.system.shouldUseWatermarker) {
|
||||
imageOutput = addWatermarker(g, imageOutput);
|
||||
}
|
||||
|
||||
g.setMetadataReceivingNode(imageOutput);
|
||||
return g.getGraph();
|
||||
};
|
||||
@@ -1,205 +0,0 @@
|
||||
import type { RootState } from 'app/store/store';
|
||||
import type { CanvasManager } from 'features/controlLayers/konva/CanvasManager';
|
||||
import { fetchModelConfigWithTypeGuard } from 'features/metadata/util/modelFetchingHelpers';
|
||||
import {
|
||||
CLIP_SKIP,
|
||||
CONTROL_LAYERS_GRAPH,
|
||||
DENOISE_LATENTS,
|
||||
LATENTS_TO_IMAGE,
|
||||
MAIN_MODEL_LOADER,
|
||||
NEGATIVE_CONDITIONING,
|
||||
NEGATIVE_CONDITIONING_COLLECT,
|
||||
NOISE,
|
||||
POSITIVE_CONDITIONING,
|
||||
POSITIVE_CONDITIONING_COLLECT,
|
||||
VAE_LOADER,
|
||||
} from 'features/nodes/util/graph/constants';
|
||||
import { addControlAdapters } from 'features/nodes/util/graph/generation/addControlAdapters';
|
||||
// import { addHRF } from 'features/nodes/util/graph/generation/addHRF';
|
||||
import { addIPAdapters } from 'features/nodes/util/graph/generation/addIPAdapters';
|
||||
import { addLoRAs } from 'features/nodes/util/graph/generation/addLoRAs';
|
||||
import { addNSFWChecker } from 'features/nodes/util/graph/generation/addNSFWChecker';
|
||||
import { addSeamless } from 'features/nodes/util/graph/generation/addSeamless';
|
||||
import { addWatermarker } from 'features/nodes/util/graph/generation/addWatermarker';
|
||||
import type { GraphType } from 'features/nodes/util/graph/generation/Graph';
|
||||
import { Graph } from 'features/nodes/util/graph/generation/Graph';
|
||||
import { getBoardField, getPresetModifiedPrompts , getSizes } from 'features/nodes/util/graph/graphBuilderUtils';
|
||||
import { isEqual } from 'lodash-es';
|
||||
import type { Invocation } from 'services/api/types';
|
||||
import { isNonRefinerMainModelConfig } from 'services/api/types';
|
||||
import { assert } from 'tsafe';
|
||||
|
||||
import { addRegions } from './addRegions';
|
||||
|
||||
export const buildTextToImageSD1SD2Graph = async (state: RootState, manager: CanvasManager): Promise<GraphType> => {
|
||||
const { bbox, params } = state.canvasV2;
|
||||
|
||||
const {
|
||||
model,
|
||||
cfgScale: cfg_scale,
|
||||
cfgRescaleMultiplier: cfg_rescale_multiplier,
|
||||
scheduler,
|
||||
steps,
|
||||
clipSkip: skipped_layers,
|
||||
shouldUseCpuNoise,
|
||||
vaePrecision,
|
||||
seed,
|
||||
vae,
|
||||
} = params;
|
||||
|
||||
assert(model, 'No model found in state');
|
||||
|
||||
const { positivePrompt, negativePrompt } = getPresetModifiedPrompts(state);
|
||||
const { originalSize, scaledSize } = getSizes(bbox);
|
||||
|
||||
const g = new Graph(CONTROL_LAYERS_GRAPH);
|
||||
const modelLoader = g.addNode({
|
||||
type: 'main_model_loader',
|
||||
id: MAIN_MODEL_LOADER,
|
||||
model,
|
||||
});
|
||||
const clipSkip = g.addNode({
|
||||
type: 'clip_skip',
|
||||
id: CLIP_SKIP,
|
||||
skipped_layers,
|
||||
});
|
||||
const posCond = g.addNode({
|
||||
type: 'compel',
|
||||
id: POSITIVE_CONDITIONING,
|
||||
prompt: positivePrompt,
|
||||
});
|
||||
const posCondCollect = g.addNode({
|
||||
type: 'collect',
|
||||
id: POSITIVE_CONDITIONING_COLLECT,
|
||||
});
|
||||
const negCond = g.addNode({
|
||||
type: 'compel',
|
||||
id: NEGATIVE_CONDITIONING,
|
||||
prompt: negativePrompt,
|
||||
});
|
||||
const negCondCollect = g.addNode({
|
||||
type: 'collect',
|
||||
id: NEGATIVE_CONDITIONING_COLLECT,
|
||||
});
|
||||
const noise = g.addNode({
|
||||
type: 'noise',
|
||||
id: NOISE,
|
||||
seed,
|
||||
width: scaledSize.width,
|
||||
height: scaledSize.height,
|
||||
use_cpu: shouldUseCpuNoise,
|
||||
});
|
||||
const denoise = g.addNode({
|
||||
type: 'denoise_latents',
|
||||
id: DENOISE_LATENTS,
|
||||
cfg_scale,
|
||||
cfg_rescale_multiplier,
|
||||
scheduler,
|
||||
steps,
|
||||
denoising_start: 0,
|
||||
denoising_end: 1,
|
||||
});
|
||||
const l2i = g.addNode({
|
||||
type: 'l2i',
|
||||
id: LATENTS_TO_IMAGE,
|
||||
fp32: vaePrecision === 'fp32',
|
||||
board: getBoardField(state),
|
||||
// This is the terminal node and must always save to gallery.
|
||||
is_intermediate: false,
|
||||
use_cache: false,
|
||||
});
|
||||
const vaeLoader =
|
||||
vae?.base === model.base
|
||||
? g.addNode({
|
||||
type: 'vae_loader',
|
||||
id: VAE_LOADER,
|
||||
vae_model: vae,
|
||||
})
|
||||
: null;
|
||||
|
||||
let imageOutput: Invocation<'l2i'> | Invocation<'img_nsfw'> | Invocation<'img_watermark'> | Invocation<'img_resize'> =
|
||||
l2i;
|
||||
|
||||
g.addEdge(modelLoader, 'unet', denoise, 'unet');
|
||||
g.addEdge(modelLoader, 'clip', clipSkip, 'clip');
|
||||
g.addEdge(clipSkip, 'clip', posCond, 'clip');
|
||||
g.addEdge(clipSkip, 'clip', negCond, 'clip');
|
||||
g.addEdge(posCond, 'conditioning', posCondCollect, 'item');
|
||||
g.addEdge(negCond, 'conditioning', negCondCollect, 'item');
|
||||
g.addEdge(posCondCollect, 'collection', denoise, 'positive_conditioning');
|
||||
g.addEdge(negCondCollect, 'collection', denoise, 'negative_conditioning');
|
||||
g.addEdge(noise, 'noise', denoise, 'noise');
|
||||
g.addEdge(denoise, 'latents', l2i, 'latents');
|
||||
|
||||
const modelConfig = await fetchModelConfigWithTypeGuard(model.key, isNonRefinerMainModelConfig);
|
||||
assert(modelConfig.base === 'sd-1' || modelConfig.base === 'sd-2');
|
||||
|
||||
g.upsertMetadata({
|
||||
generation_mode: 'txt2img',
|
||||
cfg_scale,
|
||||
cfg_rescale_multiplier,
|
||||
width: scaledSize.width,
|
||||
height: scaledSize.height,
|
||||
positive_prompt: positivePrompt,
|
||||
negative_prompt: negativePrompt,
|
||||
model: Graph.getModelMetadataField(modelConfig),
|
||||
seed,
|
||||
steps,
|
||||
rand_device: shouldUseCpuNoise ? 'cpu' : 'cuda',
|
||||
scheduler,
|
||||
clip_skip: skipped_layers,
|
||||
vae: vae ?? undefined,
|
||||
});
|
||||
|
||||
const seamless = addSeamless(state, g, denoise, modelLoader, vaeLoader);
|
||||
|
||||
addLoRAs(state, g, denoise, modelLoader, seamless, clipSkip, posCond, negCond);
|
||||
|
||||
// We might get the VAE from the main model, custom VAE, or seamless node.
|
||||
const vaeSource = seamless ?? vaeLoader ?? modelLoader;
|
||||
g.addEdge(vaeSource, 'vae', l2i, 'vae');
|
||||
|
||||
if (!isEqual(scaledSize, originalSize)) {
|
||||
// We are using scaled bbox and need to resize the output image back to the original size.
|
||||
imageOutput = g.addNode({
|
||||
id: 'img_resize',
|
||||
type: 'img_resize',
|
||||
...originalSize,
|
||||
is_intermediate: false,
|
||||
use_cache: false,
|
||||
});
|
||||
g.addEdge(l2i, 'image', imageOutput, 'image');
|
||||
}
|
||||
|
||||
const _addedCAs = addControlAdapters(state.canvasV2.controlAdapters.entities, g, denoise, modelConfig.base);
|
||||
const _addedIPAs = addIPAdapters(state.canvasV2.ipAdapters.entities, g, denoise, modelConfig.base);
|
||||
const _addedRegions = await addRegions(
|
||||
manager,
|
||||
state.canvasV2.regions.entities,
|
||||
g,
|
||||
state.canvasV2.document,
|
||||
state.canvasV2.bbox,
|
||||
modelConfig.base,
|
||||
denoise,
|
||||
posCond,
|
||||
negCond,
|
||||
posCondCollect,
|
||||
negCondCollect
|
||||
);
|
||||
|
||||
// const isHRFAllowed = !addedLayers.some((l) => isInitialImageLayer(l) || isRegionalGuidanceLayer(l));
|
||||
// if (isHRFAllowed && state.hrf.hrfEnabled) {
|
||||
// imageOutput = addHRF(state, g, denoise, noise, l2i, vaeSource);
|
||||
// }
|
||||
|
||||
if (state.system.shouldUseNSFWChecker) {
|
||||
imageOutput = addNSFWChecker(g, imageOutput);
|
||||
}
|
||||
|
||||
if (state.system.shouldUseWatermarker) {
|
||||
imageOutput = addWatermarker(g, imageOutput);
|
||||
}
|
||||
|
||||
g.setMetadataReceivingNode(imageOutput);
|
||||
return g.getGraph();
|
||||
};
|
||||
Reference in New Issue
Block a user