Commit 7aa4d93

committed

fix: adjust timestep calculations for DDIM and TCD

On img2img, the number of steps correspond to the last precalculated sigma values, but the internal alphas_cumprod and compvis_sigmas were being computed over the entire step range. Also, tweaks the prev_timestep calculation on DDIM to better match the current timestamp (like on TCD), to avoid inconsistencies due to rounding.

1 parent 10c6501 commit 7aa4d93Copy full SHA for 7aa4d93

File tree

2 files changed

+21

-8

lines changed

denoiser.hpp
stable-diffusion.cpp

2 files changed

+21

-8

lines changed

`‎denoiser.hpp‎`

Lines changed: 12 additions & 6 deletions

Original file line number	Diff line number	Diff line change
`@@ -474,6 +474,7 @@ static void sample_k_diffusion(sample_method_t method,`
`474`	`474`	`ggml_context* work_ctx,`
`475`	`475`	`ggml_tensor* x,`
`476`	`476`	`std::vector<float> sigmas,`
	`477`	`+ int initial_step,`
`477`	`478`	`std::shared_ptr<RNG> rng,`
`478`	`479`	`float eta) {`
`479`	`480`	`size_t steps = sigmas.size() - 1;`
`@@ -1060,10 +1061,14 @@ static void sample_k_diffusion(sample_method_t method,`
`1060`	`1061`	`// x_t"`
`1061`	`1062`	`// - pred_prev_sample -> "x_t-1"`
`1062`	`1063`	`int timestep =`
`1063`		`- roundf(TIMESTEPS -`
`1064`		`- i * ((float)TIMESTEPS / steps)) - 1;`
	`1064`	`+ TIMESTEPS - 1 -`
	`1065`	`+ (int)roundf((initial_step + i) *`
	`1066`	`+ (TIMESTEPS / float(initial_step + steps)));`
`1065`	`1067`	`// 1. get previous step value (=t-1)`
`1066`		`- int prev_timestep = timestep - TIMESTEPS / steps;`
	`1068`	`+ int prev_timestep =`
	`1069`	`+ TIMESTEPS - 1 -`
	`1070`	`+ (int)roundf((initial_step + i + 1) *`
	`1071`	`+ (TIMESTEPS / float(initial_step + steps)));`
`1067`	`1072`	`// The sigma here is chosen to cause the`
`1068`	`1073`	`// CompVisDenoiser to produce t = timestep`
`1069`	`1074`	`float sigma = compvis_sigmas[timestep];`
`@@ -1236,12 +1241,13 @@ static void sample_k_diffusion(sample_method_t method,`
`1236`	`1241`	`// Analytic form for TCD timesteps`
`1237`	`1242`	`int timestep = TIMESTEPS - 1 -`
`1238`	`1243`	`(TIMESTEPS / original_steps) *`
`1239`		`- (int)floor(i * ((float)original_steps / steps));`
	`1244`	`+ (int)floor((initial_step + i) *`
	`1245`	`+ ((float)original_steps / (initial_step + steps)));`
`1240`	`1246`	`// 1. get previous step value`
`1241`	`1247`	`int prev_timestep = i >= steps - 1 ? 0 :`
`1242`	`1248`	`TIMESTEPS - 1 - (TIMESTEPS / original_steps) *`
`1243`		`- (int)floor((i + 1) *`
`1244`		`- ((float)original_steps / steps));`
	`1249`	`+ (int)floor((initial_step + i + 1) *`
	`1250`	`+ ((float)original_steps / (initial_step + steps)));`
`1245`	`1251`	`// Here timestep_s is tau_n' in Algorithm 4. The _s`
`1246`	`1252`	`// notation appears to be that from C. Lu,`
`1247`	`1253`	`// "DPM-Solver: A Fast ODE Solver for Diffusion`

`‎stable-diffusion.cpp‎`

Lines changed: 9 additions & 2 deletions

Original file line number	Diff line number	Diff line change
`@@ -798,6 +798,7 @@ class StableDiffusionGGML {`
`798`	`798`	`float eta,`
`799`	`799`	`sample_method_t method,`
`800`	`800`	`const std::vector<float>& sigmas,`
	`801`	`+ int initial_step,`
`801`	`802`	`int start_merge_step,`
`802`	`803`	`SDCondition id_cond,`
`803`	`804`	`std::vector<int> skip_layers = {},`
`@@ -991,7 +992,7 @@ class StableDiffusionGGML {`
`991`	`992`	`return denoised;`
`992`	`993`	`};`
`993`	`994`
`994`		`- sample_k_diffusion(method, denoise, work_ctx, x, sigmas, rng, eta);`
	`995`	`+ sample_k_diffusion(method, denoise, work_ctx, x, sigmas, initial_step, rng, eta);`
`995`	`996`
`996`	`997`	`x = denoiser->inverse_noise_scaling(sigmas[sigmas.size() - 1], x);`
`997`	`998`
`@@ -1202,6 +1203,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,`
`1202`	`1203`	`int height,`
`1203`	`1204`	`enum sample_method_t sample_method,`
`1204`	`1205`	`const std::vector<float>& sigmas,`
	`1206`	`+ int initial_step,`
`1205`	`1207`	`int64_t seed,`
`1206`	`1208`	`int batch_count,`
`1207`	`1209`	`const sd_image_t* control_cond,`
`@@ -1464,6 +1466,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx,`
`1464`	`1466`	`eta,`
`1465`	`1467`	`sample_method,`
`1466`	`1468`	`sigmas,`
	`1469`	`+ initial_step,`
`1467`	`1470`	`start_merge_step,`
`1468`	`1471`	`id_cond,`
`1469`	`1472`	`skip_layers,`
`@@ -1611,6 +1614,7 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,`
`1611`	`1614`	`height,`
`1612`	`1615`	`sample_method,`
`1613`	`1616`	`sigmas,`
	`1617`	`+ 0,`
`1614`	`1618`	`seed,`
`1615`	`1619`	`batch_count,`
`1616`	`1620`	`control_cond,`
`@@ -1775,8 +1779,9 @@ sd_image_t* img2img(sd_ctx_t* sd_ctx,`
`1775`	`1779`	`if (t_enc == sample_steps)`
`1776`	`1780`	`t_enc--;`
`1777`	`1781`	`LOG_INFO("target t_enc is %zu steps", t_enc);`
	`1782`	`+ int initial_step = sample_steps - t_enc - 1;`
`1778`	`1783`	`std::vector<float> sigma_sched;`
`1779`		`- sigma_sched.assign(sigmas.begin() + sample_steps - t_enc - 1, sigmas.end());`
	`1784`	`+ sigma_sched.assign(sigmas.begin() + initial_step, sigmas.end());`
`1780`	`1785`
`1781`	`1786`	`sd_image_t* result_images = generate_image(sd_ctx,`
`1782`	`1787`	`work_ctx,`
`@@ -1791,6 +1796,7 @@ sd_image_t* img2img(sd_ctx_t* sd_ctx,`
`1791`	`1796`	`height,`
`1792`	`1797`	`sample_method,`
`1793`	`1798`	`sigma_sched,`
	`1799`	`+ initial_step,`
`1794`	`1800`	`seed,`
`1795`	`1801`	`batch_count,`
`1796`	`1802`	`control_cond,`
`@@ -1903,6 +1909,7 @@ SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,`
`1903`	`1909`	`0.f,`
`1904`	`1910`	`sample_method,`
`1905`	`1911`	`sigmas,`
	`1912`	`+ 0,`
`1906`	`1913`	`-1,`
`1907`	`1914`	`SDCondition(NULL, NULL, NULL));`
`1908`	`1915`

0 commit comments

Comments

(0)

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Commit 7aa4d93

File tree

2 files changed

2 files changed

`‎denoiser.hpp‎`

`‎stable-diffusion.cpp‎`

0 commit comments