Docs: fold the quality notes into BRIDGES as neutral reference, drop GUIDE.md

Keyframes, H3 prompts from the timeline and depth are documented in BRIDGES.md;
launch-flag advice is hardware-neutral.

Author: chillithebillis

CHANGELOG.md

@@ -27,7 +27,7 @@ Difforum becomes a direction layer: a visual timeline drives any renderer.
 - **Look Mix**: detail transfer, colour, flicker cuts or crossfade from a feedback pass onto
   an H3 / LTX render. New looks: disco_diffusion, vqgan_clip, flicker_experimental.
 - Template **12 · long shot with key moments** (30 s, installations).
-- **Hover help on every input** (one registry, `nodes/tooltips.py`) and a quality guide (`docs/GUIDE.md`).
+- **Hover help on every input** (one registry, `nodes/tooltips.py`, also in docs/NODES.md).
 - **H3 structured prompts**: Camera → Prompt and H3 Shot write the whole Director timeline (look,
   scenes, camera in H3 vocabulary, key events, soundscape, music) in MiniMax H3's native format,
   with the I2VA / FL2VA alignment lines or the full-reference sections.
@@ -54,8 +54,8 @@ Difforum becomes a direction layer: a visual timeline drives any renderer.
 - One shared feedback engine behind the Feedback Sampler, Live Sampler and Storyboard.
 - Depth follows the image in 3D, and pseudo-3D replaces the silent freeze without depth.
 - Steps scale with the energy, like Deforum (about 2x faster); the run report shows
-  seconds per frame and warns when ComfyUI was launched with `--lowvram` /
-  `--disable-smart-memory`, which reload the model every frame.
+  seconds per frame and warns when low-VRAM launch flags make ComfyUI reload the model
+  every frame.
 - Cadence crossfades between keys. Revealed areas are repainted with extra noise.
 - Colour anchoring per scene. Lazy prompt travel. Ring buffer in the Live Sampler.
 - Deterministic 3D z-buffer on CUDA / MPS / CPU.

README.md

@@ -155,8 +155,7 @@ Deforum hard to use:
 
 | | |
 |---|---|
-| [Quality guide](docs/GUIDE.md) | The dials that matter, keyframe quality, depth, AE cameras, H3 prompts |
-| [Bridges](docs/BRIDGES.md) | MiniMax H3, LTX-2, After Effects, Blender |
+| [Bridges](docs/BRIDGES.md) | MiniMax H3, LTX-2, keyframes, H3 prompts, depth, After Effects, Blender |
 | [Node reference](docs/NODES.md) | Every input and output |
 | [Performance](docs/PERFORMANCE.md) | Speed vs. quality, Apple Silicon |
 | [Models](docs/MODELS.md) | What works well, and the settings |

docs/BRIDGES.md

@@ -168,3 +168,51 @@ one picture per key (or takes times such as `0, 4s, 9.5s`) and outputs
 The **Animatic** is the previz of the whole shot (low resolution, a few seconds
 even for a minute of footage). Press **Previz only** on the Director to mute every
 render output, queue, check, then press it again to render.
+
+## Keyframes for video models
+
+The guide keyframes are what H3 / LTX copy: their sharpness, palette and texture.
+
+**How many.** The Director's **Keys** track marks exact moments (with Keyframe
+Images). **Keyframes** samples a clip every `every_seconds` (or at explicit
+`indices` such as `0, 48, 96, -1`) on the model's latent grid. **H3 Guides**
+keeps up to `max_guides` of them: first, last, and the best spread in time. More
+guides follow the frames closely; fewer leave more to the model.
+
+**Quality.** Guide Frames warps one image along the camera, so keyframes soften as
+the camera travels and the uncovered areas are empty. The H3 templates run three
+switchable blocks before H3 Guides:
+
+| block | what it does |
+|---|---|
+| **Depth (Depth Anything 3)** | core depth estimation on the anchor, so 3d moves have real parallax |
+| **Fill Reveal (AI)** | an image model paints the uncovered areas (inpainting checkpoints seam best) |
+| **Keyframe Polish** | Restyle `clean restyle` at denoise ~0.35 re-paints each keyframe: detail back, composition kept |
+
+Consistent keyframes (same model, prompt and seed) interpolate better than a mix of
+styles. Hand-made stills on the Keys track beat warped frames for important moments.
+
+## H3 prompts from the timeline
+
+Camera → Prompt (`format = H3 structured`) and H3 Shot (`prompt_style = H3
+structured`) write the Director timeline in MiniMax H3's native prompt format:
+
+| timeline | prompt |
+|---|---|
+| look | style that opens `[Shot 1]` |
+| scenes | what is on screen; a transformation in one take, or a new `[Shot N]` at its cut time with `cuts` on |
+| camera blocks | H3 camera vocabulary (push in, truck, arc, pedestal...) with amplitude and speed |
+| key labels | the event at that moment |
+| `soundscape` / `music` | `overall_soundscape` / `non_diegetic_music` |
+
+`h3_mode` adds the I2VA or FL2VA alignment line, or writes the full-reference
+sections for ref2va with H3 Guides. Scene prompts and key labels work best as
+things that can be seen or heard.
+
+## Depth
+
+Set the Director to **3d** and connect a depth map (white = near) to Guide Frames
+or the Feedback Sampler. Templates 03, 08 and 10 use the core Depth Anything 3
+nodes (`depth_anything_3_mono_large` in `models/geometry_estimation`, `v2_style`
+render). The Feedback Sampler also outputs the tracked depth per frame for
+compositing. Lower `translation_scale` if the parallax is too strong.

docs/GUIDE.md

@@ -20,11 +20,10 @@ Only the first row is worth optimising. The rest is noise.
 
 ## Before anything: launch flags and step scaling
 
-- **Launch flags.** `--lowvram`, `--novram` and `--disable-smart-memory` (useful for
-  MiniMax H3 / LTX on 24 GB) make ComfyUI re-stage the image model for *every*
-  frame of a feedback render. SDXL fits easily on a 24 GB card: start ComfyUI
-  without these flags for Feedback / Live renders. The run report warns when they
-  are on.
+- **Launch flags.** `--lowvram`, `--novram` and `--disable-smart-memory` make
+  ComfyUI re-stage the image model for *every* frame of a feedback render. If the
+  image model fits in VRAM, run Feedback / Live renders without them. The run
+  report warns when they are on.
 - **Steps scale with the energy** (`step_scaling = by energy`, the default): a
   frame at denoise 0.5 runs 10 of 20 steps, as in Deforum. `fixed` runs all steps.
 

docs/PERFORMANCE.md

@@ -56,7 +56,7 @@
     "Node name for S&R": "MarkdownNote"
    },
    "widgets_values": [
-    "## Loop without a crossfade\n\nThe Camera path is made periodic over 120 frames and rendered for 3 laps; the feedback settles onto its cycle and **Loop** keeps the last lap. For projections that run for hours.\n\n**Speed:** cadence 2 and steps x energy (10 of 20 at 0.5). Faster: a DMD2 / Lightning LoRA (steps 4-6, cfg 1-2) or `long_edge` 512. Launch ComfyUI without `--lowvram` / `--disable-smart-memory` for feedback renders: they reload the model every frame."
+    "## Loop without a crossfade\n\nThe Camera path is made periodic over 120 frames and rendered for 3 laps; the feedback settles onto its cycle and **Loop** keeps the last lap. For projections that run for hours.\n\n**Speed:** cadence 2 and steps x energy (10 of 20 at 0.5). Faster: a DMD2 / Lightning LoRA (steps 4-6, cfg 1-2) or `long_edge` 512. Low-VRAM launch flags (`--lowvram`, `--disable-smart-memory`) reload the model every frame."
    ]
   },
   {

example_workflows/06_seamless_loop.json

@@ -115,7 +115,7 @@ def launch_warning() -> str:
     if not bad:
         return ""
     msg = (f"launched with {' '.join(bad)}: the image model is re-loaded for every frame. "
-           "For Feedback / Live renders start ComfyUI without these flags (keep them for H3 / LTX).")
+           "If the image model fits in VRAM, run Feedback / Live renders without these flags.")
     if not _WARNED:
         _WARNED.append(1)
         print(f"[Difforum] {msg}")

nodes/render.py

@@ -609,9 +609,8 @@ def wf_loop():
     control(g, "## Loop without a crossfade\n\nThe Camera path is made periodic over 120 frames and "
             "rendered for 3 laps; the feedback settles onto its cycle and **Loop** keeps the last lap. "
             "For projections that run for hours.\n\n**Speed:** cadence 2 and steps x energy (10 of 20 at "
-            "0.5). Faster: a DMD2 / Lightning LoRA (steps 4-6, cfg 1-2) or `long_edge` 512. Launch ComfyUI "
-            "without `--lowvram` / `--disable-smart-memory` for feedback renders: they reload the model "
-            "every frame.")
+            "0.5). Faster: a DMD2 / Lightning LoRA (steps 4-6, cfg 1-2) or `long_edge` 512. Low-VRAM "
+            "launch flags (`--lowvram`, `--disable-smart-memory`) reload the model every frame.")
     with g.block(B_DIRECT, C_DIRECT, col=1, row=0):
         s = g.add("Difforum_Setup", size=(320, 300), duration_mode="frames", duration=360.0)
         cam = g.add("Difforum_Camera", size=(380, 360), keys="0: orbit_right 1.0 0.8\n60: spiral 1.2 1.0",