CHANGELOG: Veda, Latent upscaler, Attention explorer sound toggle and LoRA stage are tested on a GPU

CHANGELOG.md

@@ -24,16 +24,16 @@
 - **Prompt enhancer: Dynamic style.** A Style choice in the Enhance panel (default: unchanged). Dynamic keeps every rule of the default but asks for a concrete action beat, a change over time and secondary motion in each shot. Instructions you type still override it.
 - **Shortcuts: find, replace and common phrases.** The ⋯ menu in Shortcuts has Find and replace (search the text of every shortcut, preview each change, match case or not, whole words only or inside words too, then change them all in one go, or remove the phrase) and Common phrases (words and phrases of any length used in more than one shortcut, most shared first, with counts; pick one to correct it everywhere). Triggers are never touched.
 
-- **Sparse attention method "veda"** (H3 loader, SLA on, method veda; Experimental, UNTESTED on a GPU). A small learned predictor (Veda, arXiv 2605.30325) picks which 128-token tiles of the attention map matter and an INT8 Triton kernel computes only those, at the predictor's own trained budget (90% for text-to-video). It needs CUDA and Triton and a predictor file in ComfyUI/models/veda, chosen in "veda_predictor" (auto takes the text-to-video file, else the first). Works with LoRA (it sees q/k/v after the LoRA), MXFP8 (weights only: q/k/v stay bf16), Two stages (a size the predictor was not trained for uses the nearest tile plan, and the log says so), dense steps and the attention backend, which keeps the text refiner, masked calls and short sequences. sla_sparsity, the rising share and the references choice are not read in this first version. When Veda cannot run, the status says why and the chosen backend is installed on its own. One log line per run in the SLA line's wording (calls, tile pairs kept, asked vs real sparsity, share of full attention computed, steps at full attention). The Triton kernel is checked against an exact reference on the GPU before first use; the torch half (tiles, predictor, selection, the override on a real tiny H3) is tested on CPU.
+- **Sparse attention method "veda"** (H3 loader, SLA on, method veda; Experimental, tested on a GPU). A small learned predictor (Veda, arXiv 2605.30325) picks which 128-token tiles of the attention map matter and an INT8 Triton kernel computes only those, at the predictor's own trained budget (90% for text-to-video). It needs CUDA and Triton and a predictor file in ComfyUI/models/veda, chosen in "veda_predictor" (auto takes the text-to-video file, else the first). Works with LoRA (it sees q/k/v after the LoRA), MXFP8 (weights only: q/k/v stay bf16), Two stages (a size the predictor was not trained for uses the nearest tile plan, and the log says so), dense steps and the attention backend, which keeps the text refiner, masked calls and short sequences. sla_sparsity, the rising share and the references choice are not read in this first version. When Veda cannot run, the status says why and the chosen backend is installed on its own. One log line per run in the SLA line's wording (calls, tile pairs kept, asked vs real sparsity, share of full attention computed, steps at full attention). The Triton kernel is checked against an exact reference on the GPU before first use; the torch half (tiles, predictor, selection, the override on a real tiny H3) is tested on CPU.
 - **Veda predictor download** (Settings > Models, shown only for method veda): a button per official file that is not in models/veda yet, directly under the predictor selector; nothing once both are there (text-to-video, reference-to-video) fetches it straight from Hugging Face on your click, with the MiniMax H3 Community License named and linked beside it. FunUI does not host or bundle the file. The download streams to a `.part` file and is kept only when its size and checksum match; taggers and this share one downloader (`core/fetch.py`).
 - **Run meter** (Settings > Engine > Sampling, off by default). Logs how grainy the picture is, across it and over time, and how much it still changes at each step, then compares with the previous run of the same seed. It adds no model call, so flipping one setting and rendering again shows what that setting did.
 - **Post settings for the finished video** (Settings > Engine > Post, all off or neutral by default): Smooth jagged edges (Light or Strong; FXAA-style, only strongly contrasting edge pixels are softened), Auto contrast, Auto brightness and Colour strength. Contrast and brightness are measured once from 16 frames spread over the clip and applied as one fixed curve, so nothing flickers from frame to frame. They run on the frames before encoding, so NVENC, its CPU fallback and every codec get the same picture. Save Video reads them from the pipeline's settings object (older saved pipelines are linked to it automatically). Tried on the CPU only.
-- **LoRA Loader: "Apply to stage"** (First, Second or Both, per slot). Shown only while the Two stages module is present and not switched off for the project; the feature itself may stay off in the sampler settings. A LoRA set to First works only in the smaller first stage, one set to Second only in the full-size steps; Two stages hands each stage the model with its own LoRAs, and ComfyUI re-patches the weights between the stages. Only the model changes for First or Second (the text encoder is encoded before sampling, so its strength is not used). When Two stages is off, or a run cannot be cut in two, those LoRAs are not applied and the Log says so. Both is the old behaviour. Tried on a small model on the CPU with every feature switched on; not on a GPU yet.
+- **LoRA Loader: "Apply to stage"** (First, Second or Both, per slot). Shown only while the Two stages module is present and not switched off for the project; the feature itself may stay off in the sampler settings. A LoRA set to First works only in the smaller first stage, one set to Second only in the full-size steps; Two stages hands each stage the model with its own LoRAs, and ComfyUI re-patches the weights between the stages. Only the model changes for First or Second (the text encoder is encoded before sampling, so its strength is not used). When Two stages is off, or a run cannot be cut in two, those LoRAs are not applied and the Log says so. Both is the old behaviour. Tried on a small model on the CPU with every feature switched on; also tested on a GPU.
 - **LoRA Loader reads LoRAs written for the Diffusers layout of H3** (separate q, k and v, `ff.net`, `transformer_blocks`), such as NVIDIA's Open-H3-Regenerate-2K refiner, which used to report "matched nothing". The names are translated, q, k and v are aimed at their rows of H3's fused attention weight, and the halves of the gated MLP are put in H3's order (its layout stores them the other way round, so without that the LoRA loads and quietly does the wrong thing). Strength 1.0, as the author says it is already folded in.
-- **Attention explorer: "Change the sound too"** (Settings > Engine > Sampling). Turn it off and only the picture's attention changes: the sound's rows keep the model's own attention and the sound is no longer one of the clip regions. On models that mix picture and sound in one sequence, Prompt focus stays off while the sound is kept as is. On by default, so nothing changes until you switch it. Untested on a GPU.
+- **Attention explorer: "Change the sound too"** (Settings > Engine > Sampling). Turn it off and only the picture's attention changes: the sound's rows keep the model's own attention and the sound is no longer one of the clip regions. On models that mix picture and sound in one sequence, Prompt focus stays off while the sound is kept as is. On by default, so nothing changes until you switch it.
 - **LoRA Loader: up to 8 LoRAs in one node.** Slots 2 to 8 add a file and their own model and text encoder strengths; they apply in order after slot 1, and a slot left on None is skipped. Slot 1 keeps its old inputs, so saved pipelines load unchanged, and chaining several loaders still works.
 - **Two stages: the first stage keeps the shape of the full picture.** Each side was rounded on its own, which bent the shape by up to 6 percent (a 18x31 latent got a 12x22 first stage), so the layout was decided in a squashed picture and then stretched. The grid is now chosen so width over height matches the full one (within a few percent, in even steps).
-- **Two stages: Latent upscaler as an option** (Settings > Engine > Sampling, "How the hand-over is enlarged"). Bicubic stays the default; plain choices Mitchell, Catmull-Rom, Lanczos, Bislerp, Bilinear, Trilinear (same as Bilinear here, time is never resampled), Area and Nearest-exact are there too (Lanczos is done on the numbers, not through 8-bit pictures as ComfyUI's would, which a latent cannot survive). Or pick Latent upscaler and a file from models/latent_upscale_models (H3's convolution upscaler) to enlarge the first stage's clean guess with the learned network instead. It runs at the ratio that reaches the full size (it works in even steps, then a bicubic trim fits the exact grid), once per run, on a small latent; the file is loaded once and kept while the upscaler is the chosen method, and leaves memory when you switch to another method or turn Two stages off. A missing or unreadable file falls back to bicubic with a note. Untested on a GPU.
+- **Two stages: Latent upscaler as an option** (Settings > Engine > Sampling, "How the hand-over is enlarged"). Bicubic stays the default; plain choices Mitchell, Catmull-Rom, Lanczos, Bislerp, Bilinear, Trilinear (same as Bilinear here, time is never resampled), Area and Nearest-exact are there too (Lanczos is done on the numbers, not through 8-bit pictures as ComfyUI's would, which a latent cannot survive). Or pick Latent upscaler and a file from models/latent_upscale_models (H3's convolution upscaler) to enlarge the first stage's clean guess with the learned network instead. It runs at the ratio that reaches the full size (it works in even steps, then a bicubic trim fits the exact grid), once per run, on a small latent; the file is loaded once and kept while the upscaler is the chosen method, and leaves memory when you switch to another method or turn Two stages off. A missing or unreadable file falls back to bicubic with a note.
 - **Upscale no longer makes the next generation reload the models.** It ran as an extra ComfyUI prompt without the workflow's loaders, and ComfyUI's cache lets go of loaders a prompt does not use; it now runs as a background job of FunUI, beside ComfyUI's queue (Interrupt in ComfyUI no longer stops it).
 - **Interpolate** (module of its own, add-on **NVIDIA RTX Frame Generation**): adds in-between frames to a finished render with NVIDIA's RTX frame generation (2 to 4 times the frames; the length and the sound stay, 24 fps becomes 48 fps). Settings > Editor > Interpolate: Never, a button beside the rating, or always when a render finishes. Quality Fast, Balanced or Best, and 10-bit processing. It runs as a background job of FunUI, not as a ComfyUI prompt, so it does not make ComfyUI drop the workflow's loaded models (an extra prompt without loaders does exactly that, and the next generation reloads them); the NVIDIA model also stays loaded between presses. Works beside Upscale in either order and never runs twice on the same file. Needs an RTX card and nvidia-vfx 0.2 or newer: switching the add-on on moves an older nvidia-vfx up (nothing else is touched), and Super Resolution never moves a newer one back. ComfyUI's own RIFE/FILM interpolation nodes are a different thing and are not used. Works on a rental GPU (a few seconds for a clip).
 - **Project settings > Resolution is the** (H3, LTX image to video): what Width and Height mean once the scene picture sets the shape. Short side (the smaller number is the clip's shorter side), Long side (the larger is its longer side) or Pixel count (the old rule: about Width x Height pixels, which turned 1024x1024 into 1376x768 for a wide picture). New projects start on Short side; projects saved earlier keep Pixel count, so nothing changes size by itself. The H3 preset no longer needs the KJNodes ImageTransformKJ node for this (saved H3 pipelines are switched over on load).