I am too stupid to get ltx-2.3 to produce smooth I2V-videos at larger resolutions....

[martinbu69]

Hi all,

I don't know if this is apropriate to ask here, but I am trying to create I2V videos with ltx-2.3... For a couple of reasons, I wrote/vibe-coded my own python wrapper that auto-generates a python interface around sd.cpp, which works generally nicely and provides an always up-to-date interface. Z-image(-turbo), Flux.2(-dev/9b-klein), ideogram4, ... they all work like a charm. But with ltx-2.3 (/w or /wo distilled LoRA) I am konstantly failing in generating videos. When using smaller resolutions, I keep getting motion, but it starts melting in 90% of the cases for 81, or 121 frames long videos. When using larger resolutions (>= width=960; height=544) pretty much 100% of the time the motion is completely frozen. (while sometimes the audio-track is fine)

I can only think of using parameters in the wrong combinations....
The (flattend) video-parameters are:

---

Video-Parameter:
Bn_compute_blocks: c_int
Fn_compute_blocks: c_int
circular_x: c_bool
circular_y: c_bool
clip_skip: c_int
control_frames: LP_sd_image_t
control_frames_size: c_int
denoising_strength: c_float
end_image: sd_image_t
end_percent: c_float
error_decay_rate: c_float
extra_tiling_args: c_char_p
fps: c_int
height: c_int
high_noise_sample_params__custom_sigmas: LP_c_float
high_noise_sample_params__custom_sigmas_count: c_int
high_noise_sample_params__eta: c_float
high_noise_sample_params__extra_sample_args: c_char_p
high_noise_sample_params__flow_shift: c_float
high_noise_sample_params__guidance__distilled_guidance: c_float
high_noise_sample_params__guidance__img_cfg: c_float
high_noise_sample_params__guidance__slg__layer_count: c_ulong
high_noise_sample_params__guidance__slg__layer_end: c_float
high_noise_sample_params__guidance__slg__layer_start: c_float
high_noise_sample_params__guidance__slg__layers: LP_c_int
high_noise_sample_params__guidance__slg__scale: c_float
high_noise_sample_params__guidance__txt_cfg: c_float
high_noise_sample_params__sample_method: c_int
high_noise_sample_params__sample_steps: c_int
high_noise_sample_params__scheduler: c_int
high_noise_sample_params__shifted_timestep: c_int
hires__custom_sigmas: LP_c_float
hires__custom_sigmas_count: c_int
hires__enabled: c_bool
hires__scale: c_float
init_image: sd_image_t
lora_count: c_uint
loras: LP_sd_lora_t
max_cached_steps: c_int
max_continuous_cached_steps: c_int
max_warmup_steps: c_int
mode: c_int
model_path: c_char_p
moe_boundary: c_float
negative_prompt: c_char_p
prompt: c_char_p
rel_size_x: c_float
rel_size_y: c_float
reset_error_on_compute: c_bool
residual_diff_threshold: c_float
reuse_threshold: c_float
sample_params__custom_sigmas: LP_c_float
sample_params__custom_sigmas_count: c_int
sample_params__eta: c_float
sample_params__extra_sample_args: c_char_p
sample_params__flow_shift: c_float
sample_params__guidance__distilled_guidance: c_float
sample_params__guidance__img_cfg: c_float
sample_params__guidance__slg__layer_count: c_ulong
sample_params__guidance__slg__layer_end: c_float
sample_params__guidance__slg__layer_start: c_float
sample_params__guidance__slg__layers: LP_c_int
sample_params__guidance__slg__scale: c_float
sample_params__guidance__txt_cfg: c_float
sample_params__sample_method: c_int
sample_params__sample_steps: c_int
sample_params__scheduler: c_int
sample_params__shifted_timestep: c_int
scm_mask: c_char_p
scm_policy_dynamic: c_bool
seed: c_long
spectrum_flex_window: c_float
spectrum_lam: c_float
spectrum_m: c_int
spectrum_stop_percent: c_float
spectrum_w: c_float
spectrum_warmup_steps: c_int
spectrum_window_size: c_int
start_percent: c_float
steps: c_int
strength: c_float
target_height: c_int
target_overlap: c_float
target_width: c_int
taylorseer_n_derivatives: c_int
taylorseer_skip_interval: c_int
temporal_tiling: c_bool
tile_size_x: c_int
tile_size_y: c_int
upscale_tile_size: c_int
upscaler: c_int
use_relative_threshold: c_bool
vace_strength: c_float
vae_tiling_params__enabled: c_bool
video_frames: c_int
width: c_int

---

Any hints??

[vectornguyen76]

Start with an A/B against sd-cli, which tells you whether the problem is the model settings or your wrapper. Run the I2V command from docs/ltx2.md#L67-L71 (1280x720, 33 frames, euler) with -v. With -v the CLI prints its full generation params (main.cpp#L696-L698). If the CLI moves at that size and your wrapper freezes, diff those params against what you pass.

One thing in your flattened list is easy to trip over. Only names that collide get a prefix. The unprefixed steps, model_path, upscaler, denoising_strength, target_width/target_height and upscale_tile_size all belong to hires (sd_hires_params_t), not to the video params (stable-diffusion.h#L403-L415). The sampler step count is sample_params__sample_steps, and the diffusion model path lives in the context params.

Also, if your generator doesn't call sd_vid_gen_params_init() for the video struct, a zero-filled ctypes struct isn't the default. flow_shift, eta and img_cfg use INFINITY, and sample_method/scheduler use *_COUNT, to mean "model default". A 0 in flow_shift, for example, is applied as the actual shift (stable-diffusion.cpp#L458-L475, #L614-L642).

Starting values from the server's recipes (defaults.md#L62-L63):

• LTX-2.3 dev: CFG 3, 20 steps, 768x512
• LTX-2.3 distilled: CFG 1, 8 steps, 768x512

With the distilled LoRA on dev I'd try the distilled values first. I haven't verified that combo.

For bigger output, generate at the small size and let the LTX spatial latent upscaler do the x2 (docs/ltx2.md#L91-L111). Set hires.enabled = true, hires.upscaler = SD_HIRES_UPSCALER_MODEL, hires.model_path to the ltx-2.3-spatial-upscaler-x2-1.1.safetensors path, and hires.steps (the docs use 4). Width and height are the pre-upscale size. Video hires rejects any other upscaler or an empty model_path (video.cpp#L1630-L1645).