Update grain.c: grain blending using photographic-paper response. Fixes #11335 · Pull Request #1386 · darktable-org/darktable

ghost · 2016-12-15T09:26:49Z

A sigmoidal function is used to model a the exposure-density characteristic curve of photographic paper, and the lightness channel is used to modulate the grain accordingly to the characteristic curve. A LUT is used for computational efficiency. The main effect is a fall off of the grain in shadows and highlights accompanied by the skewing of grain distribution (for further info see the discussion on discuss.pixls.us: https://discuss.pixls.us/t/lets-improve-grain/2709/22). A "midtones bias" slider has been added, at 0% the added grain is indistinguishable from the old implementation.

…c curves A sigmoidal function is used to model a the exposure-density characteristic curve of photographic paper, and the lightness channel is used to modulate the grain accordingly to the characteristic curve. A LUT is used for computational efficiency. The main effect is a fall off of the grain in shadows and highlights accompanied by the skewing of grain distribution (for further info see the discussion on discuss.pixls.us: https://discuss.pixls.us/t/lets-improve-grain/2709/22). A "midtones bias" slider has been added, at 0% the added grain is indistinguishable from the old implementation.

LebedevRI

Looks promising, not too intrusive, but changes are needed.

LebedevRI · 2016-12-15T09:28:22Z

+#define GRAIN_LUT_PAPER_GAMMA 1.0
+
 #define CLIP(x) ((x < 0) ? 0.0 : (x > 1.0) ? 1.0 : x)
 DT_MODULE_INTROSPECTION(1, dt_iop_grain_params_t)


Change to

DT_MODULE_INTROSPECTION(2, dt_iop_grain_params_t)

Thanks for the fast comments to the code. What does DT_MODULE_INTROSPECTION do?

Most importantly, you changed the size of params struct, so old history stacks, with params struct of old size are now broken.
Bumping version (1->2) at least prevents them from being broken silently.

As for DT_MODULE_INTROSPECTION(), it is just a magic word for introspection generator to create varous functions to be able to generate/modify the params from code easily.

LebedevRI · 2016-12-15T09:32:10Z

+  float midtones_bias;
+  float grain_lut[GRAIN_LUT_SIZE * GRAIN_LUT_SIZE];
 } dt_iop_grain_data_t;



Add:

int legacy_params(dt_iop_module_t *self, const void *const old_params, const int old_version, void *new_params, const int new_version) { if(old_version == 1 && new_version == 2) { typedef struct dt_iop_grain_params_v1_t { _dt_iop_grain_channel_t channel; float scale; float strength; } dt_iop_grain_params_v1_t; dt_iop_grain_params_v1_t *o = old_params; dt_iop_grain_params_t *n = new_params; n->channel = o->channel; n->scale = o->scale; n->strength = o->strength; n->midtones_bias = 0.0; // ??? _must_ produce the same results as before this pr return 0; } return 1; }

LebedevRI · 2016-12-15T09:33:06Z

  return total;
 }

+float paper_resp(float exposure, float mb, float gp)


LebedevRI · 2016-12-15T09:33:10Z

+  return density;
+}
+
+float paper_resp_inverse(float density, float mb, float gp)


LebedevRI · 2016-12-15T09:33:15Z

+  }
+}
+
+float dt_lut_lookup_2d_1c(const float *grain_lut, const float x, const float y)


LebedevRI · 2016-12-15T09:34:47Z

+
+static void evaluate_grain_lut(float * grain_lut, const float mb)
+{
+  for(int i = 0; i < GRAIN_LUT_SIZE; i++)


maybe

#ifdef _OPENMP #pragma omp parallel for default(none) collapse(2) #endif

LebedevRI · 2016-12-15T09:37:32Z

      }

-      out[0] = in[0] + ((100.0 * (noise * (strength))) * GRAIN_LIGHTNESS_STRENGTH_SCALE);
+      out[0] = in[0] + 100 * dt_lut_lookup_2d_1c(data->grain_lut, noise * strength * GRAIN_LIGHTNESS_STRENGTH_SCALE, in[0] / 100);


100 * dt_lut_lookup_2d_1c(..., in[0] / 100)

Move that 100 into curve, to avoid dividing, and then multiplying

LebedevRI · 2016-12-15T09:38:08Z

+  dt_iop_module_t *self = (dt_iop_module_t *)user_data;
+  if(self->dt->gui->reset) return;
+  dt_iop_grain_params_t *p = (dt_iop_grain_params_t *)self->params;
+  p->midtones_bias = dt_bauhaus_slider_get(slider) / 100;


/ 100
here and everywhere, isn't that an integer division?

Also, this does not seem nice, it is better to make p->midtones_bias contain the same scale as the slider, and if needed, let the commit_params() do the scaling.

It is not an integer division. So I should add .0f.
Ok nice, I'll move the scaling of the parameter in commit_params().

LebedevRI · 2016-12-15T11:23:52Z

  module->gui_data = NULL;
  dt_iop_grain_params_t tmp
-      = (dt_iop_grain_params_t){ DT_GRAIN_CHANNEL_LIGHTNESS, 1600.0 / GRAIN_SCALE_FACTOR, 25.0 };
+      = (dt_iop_grain_params_t){ DT_GRAIN_CHANNEL_LIGHTNESS, 1600.0 / GRAIN_SCALE_FACTOR, 25.0, 0.0 };


I would guess that the new default behavior should be to produce the grain that makes use of this pr.
I.e. that default should not be 0.0.

It is definitely a good idea.

LebedevRI · 2016-12-15T11:25:04Z

+float paper_resp(float exposure, float mb, float gp)
+{
+  float density;
+  float delta = GRAIN_LUT_DELTA_MAX * exp(mb * log(GRAIN_LUT_DELTA_MIN));


$ man 3 log :

LOG(3) Linux Programmer's Manual LOG(3) NAME log, logf, logl - natural logarithmic function SYNOPSIS #include <math.h> double log(double x);

I'm not sure you intended it to actually use double versions

LebedevRI · 2016-12-15T11:26:18Z

+{
+  float density;
+  float delta = GRAIN_LUT_DELTA_MAX * exp(mb * log(GRAIN_LUT_DELTA_MIN));
+  density = (1 + 2 * delta) / (1 + exp( (4 * gp * (0.5 - exposure)) / (1 + 2 * delta) )) - delta;


All these suffix-less constants seem scary, too easy to unintentionally do int-based operation i'd say.
Append .0f

LebedevRI · 2016-12-15T11:26:49Z

+float paper_resp_inverse(float density, float mb, float gp)
+{
+  float exposure;
+  float delta = GRAIN_LUT_DELTA_MAX * exp(mb * log(GRAIN_LUT_DELTA_MIN));


Same as for paper_resp()

LebedevRI

Seems to be going in the right direction.

LebedevRI · 2016-12-16T16:23:37Z

+static float paper_resp(float exposure, float mb, float gp)
+{
+  float density;
+  float delta = GRAIN_LUT_DELTA_MAX * exp(mb * logf(GRAIN_LUT_DELTA_MIN));


Same applies to exp(), it should probably be expf()

I wonder if it's worth to use dt_fast_expf() everywhere?

LebedevRI · 2016-12-16T16:23:45Z

+static float paper_resp_inverse(float density, float mb, float gp)
+{
+  float exposure;
+  float delta = GRAIN_LUT_DELTA_MAX * exp(mb * logf(GRAIN_LUT_DELTA_MIN));


Same applies to exp(), it should probably be expf()

LebedevRI · 2016-12-16T16:24:46Z

  g_signal_connect(G_OBJECT(g->scale2), "value-changed", G_CALLBACK(strength_callback), self);
+
+  /* midtones bias */
+  g->scale3 = dt_bauhaus_slider_new_with_range(self, 0.0, 100.0, 1.0, p->midtones_bias * 100, 2);


Hm, or you could make that slider have 0..1 range.
I'm not sure whether it is better or worse than percents though.

The current strength slider is in percent, so this one should behave the same.

Makes sense.

Makes sense to me, also.

houz · 2016-12-17T11:29:25Z

+static float paper_resp(float exposure, float mb, float gp)
+{
+  float density;
+  float delta = GRAIN_LUT_DELTA_MAX * exp(mb * logf(GRAIN_LUT_DELTA_MIN));


I wonder if it's worth to use dt_fast_expf() everywhere?

houz · 2016-12-17T11:33:27Z

      }

-      out[0] = in[0] + ((100.0 * (noise * (strength))) * GRAIN_LIGHTNESS_STRENGTH_SCALE);
+      out[0] = in[0] + (100.0 * dt_lut_lookup_2d_1c(data->grain_lut, (noise * strength) * GRAIN_LIGHTNESS_STRENGTH_SCALE, in[0] / 100.0));


Why don't you push the 100.0 * into the LUT? It's not much, but would save a few cycles. Or would that give different results? I haven't really tried to understand the logic yet.

Yes, no problem, it can be moved into the LUT.

houz · 2016-12-17T11:35:12Z

+  dt_iop_module_t *self = (dt_iop_module_t *)user_data;
+  if(self->dt->gui->reset) return;
+  dt_iop_grain_params_t *p = (dt_iop_grain_params_t *)self->params;
+  p->midtones_bias = dt_bauhaus_slider_get(slider);


The other GUI callbacks do the scaling of the slider values here already.

houz · 2016-12-17T11:35:49Z

  d->channel = p->channel;
  d->scale = p->scale;
  d->strength = p->strength;
+  d->midtones_bias = p->midtones_bias / 100.0f;


Instead of scaling here you should do it in midtones_bias_callback().

No.
What about introspection?

What about it? The rest of that IOP scales in the other place already.

If for consistency with the old code in this IOP, then yes, probably better to handle it in callbacks.
But i think in new code it should be done in commit_params() if needed at all.

Edit: because displaying parameter with one range (0..100%) in gui, and with another rande in params and introspection (0..1) will probably be confusing for someone who tries to use introspection.

houz · 2016-12-17T11:37:38Z

  g_signal_connect(G_OBJECT(g->scale2), "value-changed", G_CALLBACK(strength_callback), self);
+
+  /* midtones bias */
+  g->scale3 = dt_bauhaus_slider_new_with_range(self, 0.0, 100.0, 1.0, p->midtones_bias * 100, 2);


The current strength slider is in percent, so this one should behave the same.

houz · 2016-12-17T11:38:54Z

+  dt_bauhaus_widget_set_label(g->scale3, NULL, _("midtones bias"));
+  dt_bauhaus_slider_set_format(g->scale3, "%.0f%%");
+  gtk_box_pack_start(GTK_BOX(self->widget), GTK_WIDGET(g->scale3), TRUE, TRUE, 0);
+  gtk_widget_set_tooltip_text(g->scale3, _("amount of midtones bias from the photographic paper response modeling"));


Maybe add a sentence about what low/high values do. Not critical, but who reads the manual … 😄

I moved the scaling of midtone_bias into the LUT functions, now the parameter is always a percentage. I added also `dt_fast_expf()`.

I noticed some luminance shifting using `dt_fast_expf()`, changing back to `expf()`.

ghost · 2017-01-16T16:28:37Z

What remains to be done in your opinion?

houz · 2017-01-16T16:52:54Z

I didn't try it yet, but the code looks ok to me.

TurboGit · 2017-01-17T15:04:06Z

I've tested it but I cannot see a difference when using the midtone-bias slider. I've looked at the image closely in different location. It seems that the image is not changed. The histogram moves a bit, so I suppose that there is some changes done to the image. Where should I looked? How this slider is working? Thanks.

houz · 2017-01-17T15:14:20Z

Look in black and white parts of the image. The grain should get less there.

TurboGit · 2017-01-17T15:17:21Z

Indeed! Thanks Tobias now I see it. Not a strong effect but nice.

All my notes were addressed.

houz · 2017-01-17T16:13:19Z

Thank you, I like the result much more than the old grain.

ghost · 2017-01-17T17:20:52Z

Thank you all for helping and reviewing the code.

Update grain.c: grain blending using photographic-paper response. Fixes darktable-org#11335

LebedevRI requested changes Dec 15, 2016

View reviewed changes

LebedevRI changed the title ~~Update grain.c: grain blending using photographic-paper response~~ Update grain.c: grain blending using photographic-paper response. Fixes #11335 Dec 15, 2016

LebedevRI requested changes Dec 15, 2016

View reviewed changes

LebedevRI added the feature: enhancement current features to improve label Dec 15, 2016

Update grain.c

31a278b

LebedevRI previously requested changes Dec 16, 2016

View reviewed changes

houz requested changes Dec 17, 2016

View reviewed changes

Update grain.c

b2e9fea

I moved the scaling of midtone_bias into the LUT functions, now the parameter is always a percentage. I added also `dt_fast_expf()`.

TurboGit added the incomplete pull request needing changes to be merged label Jan 2, 2017

remove dt_fast_expf()

56b3077

I noticed some luminance shifting using `dt_fast_expf()`, changing back to `expf()`.

TurboGit removed the incomplete pull request needing changes to be merged label Jan 17, 2017

houz merged commit f52fe5c into darktable-org:master Jan 17, 2017

LebedevRI modified the milestone: 2.4 Jan 21, 2017

piratenpanda pushed a commit to piratenpanda/darktable that referenced this pull request Aug 23, 2026

Merge pull request darktable-org#1386 from arctee/master

75aef4e

Update grain.c: grain blending using photographic-paper response. Fixes darktable-org#11335

Uh oh!

Conversation

ghost commented Dec 15, 2016

Uh oh!

LebedevRI left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ghost Dec 16, 2016 • edited by ghost Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LebedevRI left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LebedevRI Dec 17, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ghost Dec 16, 2016 •

edited by ghost

Loading

LebedevRI Dec 17, 2016 •

edited

Loading