feat: release V3.5.0

This commit is contained in:
ukr8b3g-cmyk
2026-08-24 14:24:25 +09:00
parent d77a467796
commit 1655d599b7
56 changed files with 12412 additions and 95 deletions
+154
View File
@@ -0,0 +1,154 @@
# V3.5 Hi-Res Fix and Second Pass workflow guide
Status: V3.5 public connection guide. The integrated Hi-Res Fix remains Experimental; the Advanced Second Pass Bridge is the accepted Issue #8 path.
## Most important correction
The accepted Second Pass scheduler values are:
```text
BasicScheduler
scheduler = simple
steps = 10
denoise = 0.35
```
This means **10 steps and denoise 0.35**. It does not mean denoise 10.
`H3 Continuum Hi-Res Fix V3.5` does not contain its own steps or denoise value. It receives the exact `SIGMAS` produced by the workflow-side `BasicScheduler`, so this connection controls how strongly and how many steps the resized latent is refined.
The values above are the GPU-accepted baseline, not a claim that they are universally optimal for every model, sampler, resolution, or prompt.
## Standard Main connection
Normal users should use the two V3.5 nodes below:
```text
H3 Continuum Sampler V3.5
│
├─ video_latents ───────┐
├─ audio_latents ───────┤
├─ assembly_plan ───────┤
└─ refine_context ──────┤
▼
H3 Continuum Hi-Res Fix V3.5
│
├─ video_latents ──→ Core Video VAE Decode
├─ audio_latents ──→ Core Audio VAE Decode
└─ updated_assembly_plan → Assemble + Seam V3.5
```
Additional connections to `H3 Continuum Hi-Res Fix V3.5`:
| Hi-Res Fix input | Connect from |
|---|---|
| `model` | The same final H3 MODEL used by First Pass |
| `clip` | The same MiniMax H3 CLIP |
| `sampler` | Second Pass `KSamplerSelect`; accepted test used `res_multistep` |
| `sigmas` | A separate Second Pass `BasicScheduler`; accepted test used `simple`, 10 steps, denoise 0.35 |
| `video_latents` | `H3 Continuum Sampler V3.5.video_latents` |
| `audio_latents` | `H3 Continuum Sampler V3.5.audio_latents` |
| `assembly_plan` | `H3 Continuum Sampler V3.5.assembly_plan` |
| `refine_context` | `H3 Continuum Sampler V3.5.refine_context` |
| `video_vae` | The same H3 Video VAE used by First Pass/Core Decode |
Accepted Main controls:
```text
enabled = true
upscale_method = lanczos
scale_by = 2.0
refine_seed = 42 in the acceptance run
```
The node derives one actual refine seed per physical group from `refine_seed`; the report shows that derived sampling seed.
When `enabled=false`, the node returns the original video/audio LATENT objects and Assembly Plan unchanged. Sampling and VAE dependencies are lazy and are not evaluated.
## First Pass scheduler versus Second Pass scheduler
There are normally two independent scheduler roles:
```text
First Pass BasicScheduler
→ H3 Continuum Sampler V3.5.sigmas
Second Pass BasicScheduler
→ H3 Continuum Hi-Res Fix V3.5.sigmas
```
The accepted GPU run used:
```text
First Pass: simple / 20 steps / denoise 1.0
Second Pass: simple / 10 steps / denoise 0.35
```
Only the Second Pass values describe the Hi-Res Fix refinement strength.
## What Hi-Res Fix does internally
```text
First Pass physical video latent
→ H3 Video VAE Decode
→ CPU pixel resize (Lanczos in the accepted run)
→ same H3 Video VAE Encode
→ context-aware physical-group Second Pass
→ refined video latent
```
Audio is not adopted from the temporary Second Pass result. The original First Pass audio LATENT objects are returned bit-exact.
The Main node never falls back to direct 24-channel latent interpolation. If the required Video VAE is missing or the VAE contract is violated, it raises an explicit error.
## Advanced external-upscaler connection
Use `H3 Continuum Second Pass V3.5` directly only when an external latent processor such as LBH is used:
```text
H3 Continuum Sampler V3.5.video_latents
→ LBH / external H3 latent processor
→ H3 Continuum Second Pass V3.5.video_latents
Second Pass BasicScheduler.SIGMAS
→ H3 Continuum Second Pass V3.5.sigmas
```
Also connect the Sampler's original `audio_latents`, `assembly_plan`, and `refine_context`, plus the same MODEL, CLIP, sampler, and Video VAE.
Do not place `H3 Continuum Hi-Res Fix V3.5` and a separate `H3 Continuum Second Pass V3.5` in series for the standard path. Hi-Res Fix already performs one Second Pass internally; chaining both would refine twice.
## Latent Resize Utility
`H3 Continuum Latent Resize V3.5` changes only the H/W of the 24-channel video latent. It does not sample and is not the standard Main backend.
Direct 2x latent interpolation to 1024/1152 square produced persistent artifacts in testing. Therefore standalone Latent Resize `nearest-exact / scale_by=2.0`, or Main Hi-Res Fix `bislerp / scale_by=2.0`, must not be described as the accepted Main 2x configuration. The accepted Main 2x configuration is the Pixel/VAE path with pixel Lanczos.
## GPU-accepted Main condition
```text
Mode: FL2VA, 1 chunk × 5 seconds
First Pass canvas: 576 × 576
Hi-Res target: 1152 × 1152
Pixel resize: Lanczos
Second Pass sampler: res_multistep
Second Pass SIGMAS: simple / 10 steps / denoise 0.35
Output latent: (1, 24, 37, 72, 72)
Physical groups: 1
Sampling passes: 1
Final media: 1152 × 1152, 120 frames, 24 fps, 5.000 seconds
```
This integrated Main output matched every decoded video and audio frame of the accepted external Decode -> Lanczos -> Encode -> Second Pass reference.
Additional GPU results:
- Hybrid FL2VA + Reference `1 x 5s` passed through the integrated 576-to-1152 Lanczos Main path. Context-aware conditioning used the captured `refine_context`, each physical group sampled once, first-pass audio was returned bit-exact, and Core Decode / Assemble produced 120 frames at 24 fps / 5.000 seconds without persistent line or plate artifacts.
- The same Hybrid/Reference input also passed through the direct Advanced Second Pass node at 576x576, confirming the public bridge node independently of the integrated wrapper.
- Main `3 x 5s` at 576-to-1152 is not accepted on the tested RTX 5060 Ti 16 GiB. The 37T group completed its 10-step Second Pass, then the terminal 77T group failed at its first inference with CUDA OOM (`15.93 GiB` device limit, `10.04 GiB` already allocated, `1.49 GiB` requested).
Still unverified:
- Main `3 x 5s` 2x on a GPU with more than 16 GiB VRAM, or a lower-memory implementation for its terminal 77T group
- Reference/Hybrid cases longer than `1 x 5s`
- Main 2x `bislerp` quality; it was not the accepted Lanczos test