File size: 13,617 Bytes
f4a41d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
import gradio as gr

from modules import scripts
import modules.shared as shared
from modules.script_callbacks import on_cfg_denoiser, remove_current_script_callbacks
import torch, math

#import torchvision.transforms.functional as TF

import ldm_patched.modules.samplers as LDM
import modules_forge.forge_sampler

#   button to spit weighted cfg to console, better: gradio lineplot for display of weights



class CFGfadeForge(scripts.Script):
    weight = 1.0
    backup_sampling_function = None

    def __init__(self):
        self.boostStep = 0.0
        self.highStep = 0.5
        self.maxScale = 1.0
        self.fadeStep = 0.5
        self.zeroStep = 1.0
        self.minScale = 0.0
        self.reinhard = 1.0
        self.rfcgmult = 1.0
        self.centreMean = False
        self.heuristic = 0
        self.hStart = 0.0

    def title(self):
        return "CFG fade"

    def show(self, is_img2img):
        # make this extension visible in both txt2img and img2img tab.
        return scripts.AlwaysVisible

    def ui(self, *args, **kwargs):
        with gr.Accordion(open=False, label=self.title()):
            with gr.Row():
                enabled    = gr.Checkbox(value=False, label='Enable modifications to CFG')
                cntrMean   = gr.Checkbox(value=False, label='centre conds to mean')
#                scaleCFGs  = gr.Checkbox(value=False, label='scale hCFG and rCFG')
            with gr.Row():
                lowCFG1   = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.1, label='CFG 1 until step')
                maxScale  = gr.Slider(minimum=1.0, maximum=4.0,  step=0.01, value=1.0, label='boost factor')
            with gr.Row():
                boostStep = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.2, label='CFG boost start step')
                minScale  = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=1.0, label='fade factor')
            with gr.Row():
                highStep  = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.4, label='full boost at step')
                heuristic = gr.Slider(minimum=0.0, maximum=16.0, step=0.5,  value=0,   label='Heuristic CFG')
            with gr.Row():
                fadeStep  = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.5, label='CFG fade start step')
                hStart    = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.0, label='... start step')
            with gr.Row():
                zeroStep  = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.7, label='full fade at step')
                reinhard  = gr.Slider(minimum=0.0, maximum=16.0, step=0.5,  value=0.0, label='Reinhard CFG')
            with gr.Row():
                highCFG1  = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.8, label='CFG 1 after step')
                rcfgmult  = gr.Slider(minimum=0.0, maximum=1.0,  step=0.01, value=0.0, label='Rescale CFG')

        self.infotext_fields = [
            (enabled, lambda d: enabled.update(value=("cfgfade_enabled" in d))),
            (cntrMean,  "cfgfade_cntrMean"),
            (boostStep, "cfgfade_boostStep"),
            (highStep,  "cfgfade_highStep"),
            (maxScale,  "cfgfade_maxScale"),
            (fadeStep,  "cfgfade_fadeStep"),
            (zeroStep,  "cfgfade_zeroStep"),
            (minScale,  "cfgfade_minScale"),
            (lowCFG1,   "cfgfade_lowCFG1"),
            (highCFG1,  "cfgfade_highCFG1"),
            (reinhard,  "cfgfade_reinhard"),
            (rcfgmult,  "cfgfade_rcfgmult"),
            (heuristic, "cfgfade_heuristic"),
            (hStart,    "cfgfade_hStart"),
        ]

        return enabled, cntrMean, boostStep, highStep, maxScale, fadeStep, zeroStep, minScale, lowCFG1, highCFG1, reinhard, rcfgmult, heuristic, hStart

#   edited from ldm_patched/modules/samplers to add cond_scaling (initial 3 lines)
    def sampling_function(model, x, timestep, uncond, cond, cond_scale, model_options={}, seed=None):
            cond_scale *= CFGfadeForge.weight
            if cond_scale < 1.0:
                cond_scale = 1.0

            edit_strength = sum((item['strength'] if 'strength' in item else 1) for item in cond)

            if math.isclose(cond_scale, 1.0) and model_options.get("disable_cfg1_optimization", False) == False:
                uncond_ = None
            else:
                uncond_ = uncond

            for fn in model_options.get("sampler_pre_cfg_function", []):
                model, cond, uncond_, x, timestep, model_options = fn(model, cond, uncond_, x, timestep, model_options)

            cond_pred, uncond_pred = LDM.calc_cond_uncond_batch(model, cond, uncond_, x, timestep, model_options)

            if "sampler_cfg_function" in model_options:
                args = {"cond": x - cond_pred, "uncond": x - uncond_pred, "cond_scale": cond_scale, "timestep": timestep, "input": x, "sigma": timestep,
                        "cond_denoised": cond_pred, "uncond_denoised": uncond_pred, "model": model, "model_options": model_options}
                cfg_result = x - model_options["sampler_cfg_function"](args)
            elif not math.isclose(edit_strength, 1.0):
                cfg_result = uncond_pred + (cond_pred - uncond_pred) * cond_scale * edit_strength
            else:
                cfg_result = uncond_pred + (cond_pred - uncond_pred) * cond_scale

            for fn in model_options.get("sampler_post_cfg_function", []):
                args = {"denoised": cfg_result, "cond": cond, "uncond": uncond, "model": model, "uncond_denoised": uncond_pred, "cond_denoised": cond_pred,
                        "sigma": timestep, "model_options": model_options, "input": x}
                cfg_result = fn(args)

            return cfg_result

    def patch(self, model):
#        sigmin = model.model.model_sampling.sigma(model.model.model_sampling.timestep(model.model.model_sampling.sigma_min))
#        sigmax = model.model.model_sampling.sigma(model.model.model_sampling.timestep(model.model.model_sampling.sigma_max))

        def sampler_cfgfade(args):
            cond = args["cond"]
            cond_scale = args["cond_scale"]

            if cond_scale == 1.0:
                return cond
            else:
                uncond = args["uncond"]

#   sometimes this scaling seems like a win, but only when heuristic/reinhard CFG is too high
#            if self.scaleCFGs == True:
#                heuristic = max(1.0, self.heuristic * CFGfadeForge.weight) if (self.heuristic > 0.0) else 0.0
#                reinhard  = max(1.0, self.reinhard  * CFGfadeForge.weight) if (self.reinhard  > 0.0) else 0.0
#            else:
            heuristic = self.heuristic
            reinhard = self.reinhard

            if self.centreMean == True:     # better after, but value here too?
                for b in range(len(cond)):
                    for c in range(4):
                        cond[b][c] -= cond[b][c].mean()
                        uncond[b][c] -= uncond[b][c].mean()

#   cond_scale weighting now applied in sampling_function, can avoid processing of uncond for performance increase

            thisStep = shared.state.sampling_step
            lastStep = shared.state.sampling_steps

#   heuristic scaling, higher hcfg acts to boost contrast/detail/sharpness; low reduces; quantile has effect, but not significant for quality IMO
            noisePrediction = cond - uncond
            if heuristic != 0.0 and heuristic != cond_scale and thisStep >= self.hStart * lastStep:
                base = uncond + cond_scale * noisePrediction
                heur = uncond + heuristic * noisePrediction

                #   center both on zero
                baseC = base - base.mean()
                heurC = heur - heur.mean()
                del base, heur

                #   calc 99.0% quartiles - doesn't seem to have value as an option
                baseQ = torch.quantile(baseC.abs(), 0.99)
                heurQ = torch.quantile(heurC.abs(), 0.99)
                del baseC, heurC

                if baseQ != 0.0 and heurQ != 0.0:
                    cond *= (baseQ / heurQ)
                    uncond *= (baseQ / heurQ)
#   end: heuristic scaling

                
#   reinhard tonemap from comfy
            noisePrediction = cond - uncond
            if reinhard != 0.0 and reinhard != cond_scale:
                multiplier = 1.0 / cond_scale * reinhard
                noise_pred_vector_magnitude = (torch.linalg.vector_norm(noisePrediction, dim=(1)) + 0.0000000001)[:,None]
                noisePrediction /= noise_pred_vector_magnitude

                mean = torch.mean(noise_pred_vector_magnitude, dim=(1,2,3), keepdim=True)
                std = torch.std(noise_pred_vector_magnitude, dim=(1,2,3), keepdim=True)
                top = (std * 3 + mean) * multiplier

                noise_pred_vector_magnitude *= (1.0 / top)
                new_magnitude = noise_pred_vector_magnitude / (noise_pred_vector_magnitude + 1.0)
                new_magnitude *= top
                cond_scale *= new_magnitude
#   end: reinhard


#   rescaleCFG - maybe should be exclusive of other effects, but why restrict?
            result = uncond + cond_scale * noisePrediction
            if self.rcfgmult != 0.0:
                ro_pos = torch.std(cond, dim=(1,2,3), keepdim=True)
                ro_cfg = torch.std(result, dim=(1,2,3), keepdim=True)

                if ro_pos != 0.0 and ro_cfg != 0.0:
                    x_rescaled = result * (ro_pos / ro_cfg)
                    result = torch.lerp (result, x_rescaled, self.rcfgmult)
                    del x_rescaled                
#   end: rescaleCFG
            del noisePrediction

            return result

        m = model.clone()
        m.set_model_sampler_cfg_function(sampler_cfgfade)
        return (m, )


    def denoiser_callback(self, params):
        lastStep = params.total_sampling_steps - 1
        thisStep = params.sampling_step
        sigma = params.sigma[0]

        lowCFG1   = self.lowCFG1   * lastStep
        highStep  = self.highStep  * lastStep
        boostStep = self.boostStep * lastStep
        highCFG1  = self.highCFG1  * lastStep
        fadeStep  = self.fadeStep  * lastStep
        zeroStep  = self.zeroStep  * lastStep

        if thisStep < lowCFG1:
            boostWeight = 0.0
        elif thisStep < boostStep:
            boostWeight = 1.0
        elif thisStep < highStep:
            boostWeight = 1.0 + (self.maxScale - 1.0) * ((thisStep - boostStep) / (highStep - boostStep))
        else:
            boostWeight = self.maxScale

        if thisStep > highCFG1:
            fadeWeight = 0.0
        else:
            if thisStep < fadeStep:
                fadeWeight = 1.0
            elif thisStep < zeroStep:
                fadeWeight = 1.0 - (thisStep - fadeStep) / (zeroStep  - fadeStep)
            else:
                fadeWeight = 0.0

            # at this point, weight is in the range 0.0->1.0
            fadeWeight *= (1.0 - self.minScale)
            fadeWeight += self.minScale
            # now it is minimum->1.0

        CFGfadeForge.weight = boostWeight * fadeWeight


    def process_before_every_sampling(self, params, *script_args, **kwargs):
        enabled = script_args[0]
        if enabled:
            unet = params.sd_model.forge_objects.unet
            unet = CFGfadeForge.patch(self, unet)[0]
            params.sd_model.forge_objects.unet = unet

        return

    def process(self, params, *script_args, **kwargs):
        enabled, cntrMean, boostStep, highStep, maxScale, fadeStep, zeroStep, minScale, lowCFG1, highCFG1, reinhard, rcfgmult, heuristic, hStart = script_args

        if not enabled:
            return

        self.centreMean = cntrMean
        self.boostStep  = boostStep
        self.highStep   = highStep
        self.maxScale   = maxScale
        self.fadeStep   = fadeStep
        self.zeroStep   = zeroStep
        self.minScale   = minScale
        self.lowCFG1    = lowCFG1
        self.highCFG1   = highCFG1
        self.reinhard   = reinhard
        self.rcfgmult   = rcfgmult
        self.heuristic  = heuristic
        self.hStart     = hStart

        # logs, could save boost start/full only if boost factor > 1
        #       similar for fade
        params.extra_generation_params.update(dict(
            cfgfade_enabled   = enabled,
            cfgfade_cntrMean  = cntrMean,
            cfgfade_boostStep = boostStep,
            cfgfade_highStep  = highStep,
            cfgfade_maxScale  = maxScale,
            cfgfade_fadeStep  = fadeStep,
            cfgfade_zeroStep  = zeroStep,
            cfgfade_minScale  = minScale,
            cfgfade_lowCFG1   = lowCFG1,
            cfgfade_highCFG1  = highCFG1,
            cfgfade_reinhard  = reinhard,
            cfgfade_rcfgmult  = rcfgmult,
            cfgfade_heuristic = heuristic,
            cfgfade_hStart    = hStart,
        ))

        #   must log the parameters before fixing minScale
        self.minScale /= self.maxScale

        on_cfg_denoiser(self.denoiser_callback)

        if CFGfadeForge.backup_sampling_function == None:
            CFGfadeForge.backup_sampling_function = modules_forge.forge_sampler.sampling_function

        modules_forge.forge_sampler.sampling_function = CFGfadeForge.sampling_function
        return

    def postprocess(self, params, processed, *args):
        enabled = args[0]
        if enabled:
            if CFGfadeForge.backup_sampling_function != None:
                modules_forge.forge_sampler.sampling_function = CFGfadeForge.backup_sampling_function

        remove_current_script_callbacks()
        return