Official Ingredients Workflow from the github
hello, i had been reading the document and none of the settings that are recommended match with the official workflow? unless im missing something i need clarification, the official workflow uses dev+distilled lora, the document mentions (STG, mode stg_v, block 29, scale 1.0) which is multimodal for the dev model, so is this supposed to be 2 staged?
I have tested with dev+detailed LORA; it didn't look nearly as good as straight detailed.
I would assume that it was trained on distilled...
Distilled:
I think it was trained on 50 frames per second. If you set it to 24, you get very strange results.
I have tested with dev+detailed LORA; it didn't look nearly as good as straight detailed.
I would assume that it was trained on distilled...
are you using 50fps 960x544?
For some reason when I do 50 fps and 14 seconds long, I get only 13 seconds of video. It meant I was losing 50 frames, and the image was terrible. So in my length I added 50 (total 751). The final video came out exceptional and was exactly 14 seconds. I kept all the settings but changed to 24 frames per second instead of 50. The result wasn't the same; this time I got 15 seconds of video and visuals and also degraded sound. Ah! What if I want 5 seconds of video at 24 frames?
For some reason when I do 50 fps and 14 seconds long, I get only 13 seconds of video. It meant I was losing 50 frames, and the image was terrible. So in my length I added 50 (total 751). The final video came out exceptional and was exactly 14 seconds. I kept all the settings but changed to 24 frames per second instead of 50. The result wasn't the same; this time I got 15 seconds of video and visuals and also degraded sound. Ah! What if I want 5 seconds of video at 24 frames?
that will be 121 frames for 24fps 5 seconds
For some reason when I do 50 fps and 14 seconds long, I get only 13 seconds of video. It meant I was losing 50 frames, and the image was terrible. So in my length I added 50 (total 751). The final video came out exceptional and was exactly 14 seconds. I kept all the settings but changed to 24 frames per second instead of 50. The result wasn't the same; this time I got 15 seconds of video and visuals and also degraded sound. Ah! What if I want 5 seconds of video at 24 frames?
that will be 121 frames for 24fps 5 seconds
How did you calculate that?
For some reason when I do 50 fps and 14 seconds long, I get only 13 seconds of video. It meant I was losing 50 frames, and the image was terrible. So in my length I added 50 (total 751). The final video came out exceptional and was exactly 14 seconds. I kept all the settings but changed to 24 frames per second instead of 50. The result wasn't the same; this time I got 15 seconds of video and visuals and also degraded sound. Ah! What if I want 5 seconds of video at 24 frames?
that will be 121 frames for 24fps 5 seconds
How did you calculate that?
I have tested with dev+detailed LORA; it didn't look nearly as good as straight detailed.
I would assume that it was trained on distilled...are you using 50fps 960x544?
1280x720 at 50fps
I have tested with dev+detailed LORA; it didn't look nearly as good as straight detailed.
I would assume that it was trained on distilled...are you using 50fps 960x544?
1280x720 at 50fps
First one is better but how do you not get that issue where like a static image plays? For me it's one of my cells that play for 2 seconds then go to the video.
Hi team, it looks like the length is working now, and the clarity is almost the same as the official template. The latent was divided in half from the original dimensions of 1280×720. It wasn’t enough to maintain the same level of clarity... Now, I need to find a way to reduce the noise and prevent faces from melting.
Here is the prompt example: Realistic cinematic Chinese historical romance style. Maintain strong reference consistency for both characters and the ancient Jiangnan village street. Do not change their faces, hairstyles, costumes, colors, accessories, embroidery patterns, body proportions, architecture, lighting, weather, or overall visual style.Continuous one-shot video, no cuts, no scene transition. The entire sequence takes place in the same ancient Jiangnan water town street shown in the reference environment. After a light rain, the narrow stone-paved alley remains wet and reflective. White walls, black tiled roofs, warm glowing lanterns under wooden eaves, hanging willow branches, drifting mist, and distant green mountains remain unchanged throughout the entire shot. The atmosphere is cold, poetic, elegant, and romantic.
The camera begins behind soft willow branches in the misty street. Raindrops fall gently from the leaves as the camera slowly moves forward along the wet stone pavement. Lantern reflections shimmer in puddles, and the quiet Jiangnan street remains visible in the background.
The woman from the reference image walks toward the camera through the same rainy alley. She has long jet-black hair styled into an elegant traditional updo, decorated with green jade and white floral ornaments. She wears matching teardrop tassel earrings and a dark green strapless mermaid-style gown with three-dimensional golden floral embroidery. Her long dress moves naturally as she walks, the fabric slightly brushing the wet stone pavement.
At first, her expression is calm and distant. As she notices the man approaching, her expression gradually softens into a gentle smile.
The man from the reference image stands further down the same street. He has long dark brown hair partially tied back with a black dragon-shaped hairpin and pearl tassel hair ornament. He wears a wine-red traditional wide-sleeved robe with subtle dark patterns, golden phoenix embroidery, and a thick black fox-fur cloak. His robe moves slightly in the cool rainy wind.
The man looks at her and smiles softly. He says:
"What a coincidence. I never expected to meet you here."
The woman slows her steps. Her cheeks become slightly red, and she gives a gentle nod. She smiles and replies:
"I didn't expect it either. Shall we take a walk together?"
The camera continues moving in one uninterrupted shot. It slowly circles around them, revealing their full side profile as they stand together on the wet stone street. The man makes a graceful inviting gesture with his hand. The woman responds with a small nod.
They begin walking side by side deeper into the same Jiangnan alley. Their sleeves and clothing move naturally with each step. Their reflections appear on the wet pavement as lantern light stretches across the ground.
The camera smoothly transitions behind them without cutting, following their footsteps from the rear. Willow leaves sway in the foreground, mist flows between the buildings, and the warm lanterns glow through the rain.
In the final moments, the camera slowly pulls farther away as the two figures walk together into the foggy distance. Their silhouettes gradually disappear at the end of the same ancient stone-paved street.
Continuous cinematic camera movement, realistic fabric motion, subtle facial expressions, gentle rain, atmospheric mist, reflections on wet ground, elegant historical romance, no cuts, no environment change, no costume change, no face change.
I directly got blurred image and distorted sound using: LTX-2.3_ICLoRA_Ingredients_Single_Stage_Distilled.json. Not having issues with Comfy's example workflow for ltx 2.3 image to video.