From 69f3c0f9c3580f18314acf861d34fb7212469689 Mon Sep 17 00:00:00 2001 From: Morpheus Date: Thu, 20 Aug 2026 20:07:50 +0000 Subject: [PATCH] Refresh daily workflow review notes --- optimized/image-to-video/README.md | 1 + .../README.md | 10 +++++++ optimized/notes/flux2klein.md | 6 +++++ optimized/notes/ltx.md | 4 +++ optimized/notes/minimax.md | 8 ++++++ .../notes/reviews/2026-08-20-daily-review.md | 27 +++++++++++++++++++ optimized/notes/zimage.md | 12 ++++++++- 7 files changed, 67 insertions(+), 1 deletion(-) create mode 100644 optimized/notes/reviews/2026-08-20-daily-review.md diff --git a/optimized/image-to-video/README.md b/optimized/image-to-video/README.md index 7579477..10e108d 100644 --- a/optimized/image-to-video/README.md +++ b/optimized/image-to-video/README.md @@ -6,5 +6,6 @@ Home for maintained category masters focused on image-to-video generation. - `ltx23-3070-simple-master` - `minimax-h3-3070-single-10s-1152x640-r2v-master` +- `minimax-h3-best-current-r2v-master-480p-3x10s` - `minimax-h3-contex-loop-3070-compact-30s` - `minimax-h3-contex-loop-3070-compact-prepend-480p-3x10s` diff --git a/optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/README.md b/optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/README.md index 60497a2..328ae0e 100644 --- a/optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/README.md +++ b/optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/README.md @@ -16,10 +16,17 @@ subject-definition prompts for continuity across the chain. - `workflow.json` +## This pass + +- Tightened the clip prompts so `` carries the master scene and `` carries the primary full-body identity, matching the current Ref2VA prompt hierarchy. +- Kept the `clip_by_clip` validation flow and the `480p` / `3 x 10s` budget intact. + ## Built from - `optimized/image-to-video/minimax-h3-extender-3070-r2v-480p-clip-by-clip/` - `models/minimax/minimax-h3-extender-ref2va/` +- `models/minimax/minimax-h3-int8-r2v-javano2608-23/` +- current MiniMax Ref2VA prompt guidance in `optimized/notes/minimax-h3-ref2va-prompt-guidelines.md` - local lessons from: - `optimized/image-to-video/minimax-h3-contex-loop-3070-compact-prepend-480p-3x10s/` - `optimized/image-to-video/minimax-h3-contex-loop-3070-compact-30s/` @@ -52,6 +59,9 @@ subject-definition prompts for continuity across the chain. - the current compact `Contex Loop` masters still define the safe local budget, and this workflow inherits that budget shape instead of pretending the source defaults are free +- the newer same-day `javawock7618` `INT8 R2V` import is the strongest current + donor for tighter subject retention, so the maintained master now spells out + the scene/subject authority order more explicitly - this is the cleanest current merge of: - better continuation workflow UX - safer `3070 / 8 GB` assumptions diff --git a/optimized/notes/flux2klein.md b/optimized/notes/flux2klein.md index 877972f..6d83ad0 100644 --- a/optimized/notes/flux2klein.md +++ b/optimized/notes/flux2klein.md @@ -19,6 +19,9 @@ - Community Hugging Face workflow worth remembering: - `models/flux2klein/community-gguf-kv-yarvixpa/` - Interesting because of KV-cache / GGUF optimization ideas, not because it is the default path +- Community multi-reference donor worth watching: + - `models/flux2klein/geceff-flux2klein9b-multiref-3image/` + - Useful as a reference-control donor, but still too heavy and opinionated to replace the official `9B fp8` edit baseline by default ## Strong current heuristics @@ -29,11 +32,13 @@ - `CFG=1` - cleaner model/text-encoder pairing - avoiding broken loader substitutions +- Keep the official `9B fp8` edit master as the default lane, then borrow from the multi-reference donor only when reference control is the real problem. ## Current weak areas - We still do not have a clean, trusted, explicitly labeled `9B fp8 outpainting` community workflow import. - Some public workflow pages mention outpainting but do not expose a trustworthy raw workflow asset. +- The multi-reference donor is promising, but it is not a clean maintainable replacement for the current edit master. ## Best current optimization direction @@ -42,6 +47,7 @@ - cleaner graph organization - low-step settings when they are clearly validated - better reference/outpaint control if portable +- Treat the `geceff` multi-reference pack as a donor for future reference-heavy variants, not as the first default promotion. ## What needs Monday proof diff --git a/optimized/notes/ltx.md b/optimized/notes/ltx.md index 4cc9280..90ca6a0 100644 --- a/optimized/notes/ltx.md +++ b/optimized/notes/ltx.md @@ -11,22 +11,26 @@ - `models/ltx/ltx23-t2v-simple-single-pass/` - `models/ltx/ltx23-first-last-frame/` - `models/ltx/ltx23-character-sheet-best-face-id/` +- `models/ltx/ltx25-multiple-subject-reference/` ## Strong current heuristics - Start with simple single-pass before trying larger consistency graphs. - Prefer split-model and extracted-model routes when the source pack says they are easier locally. - Treat first-last-frame and character-sheet workflows as second-wave tests, not the baseline. +- Keep the `LTX-2.5` multi-subject reference pack as the best current donor for heavier consistency work, but do not let it displace the simpler `LTX 2.3` master until we have proof on the `3070`. ## Current weak areas - No real proof yet for what is genuinely practical on Chris's box. - Community packs bundle lots of helper nodes, which raises import fragility. +- The newer `LTX-2.5` reference pack looks promising, but it is not yet a clean default candidate. ## Best current optimization direction - Keep `optimized/image-to-video/ltx23-3070-simple-master/` as the first `LTX` default. - Only promote first-last-frame or character-sheet variants after the simple master proves viable. +- Use the `LTX-2.5` multi-subject reference workflow as a future donor for reference-heavy jobs, not as the baseline. ## What needs Monday proof diff --git a/optimized/notes/minimax.md b/optimized/notes/minimax.md index f660347..7265afd 100644 --- a/optimized/notes/minimax.md +++ b/optimized/notes/minimax.md @@ -12,17 +12,23 @@ - `models/minimax/minimax-h3-turbo-lora-community/` - `models/minimax/minimax-h3-int8-r2v-community/` - `models/minimax/minimax-h3-extender-ref2va/` +- `models/minimax/minimax-h3-int8-r2v-javano2608-23/` +- `models/minimax/minimax-h3-turbo-gguf-i2va/` +- `models/minimax/minimax-h3-seamless-chain-v2/` ## Strong current heuristics - Prefer Turbo LoRA for the first `text-to-video` attempt. - Prefer INT8-based community workflows for `reference-to-video` exploration. - Keep attention and memory-efficiency helpers if they are already integrated and well documented. +- Treat the same-day `javawock7618` `INT8 R2V` import as the strongest current donor for local continuity and identity retention, but not as proof that the graph is light enough for a `3070`. +- Keep the GGUF `I2VA` path as an audio-aware fallback, not the default lane. ## Current weak areas - No proof yet that MiniMax is truly practical on Chris's box. - Custom-node stack is more fragile than the current `flux2klein` path. +- The full `seamless-chain` graph is interesting, but it is not a clean baseline for the first maintained master. ## Best current optimization direction @@ -30,6 +36,7 @@ - Keep the INT8 reference-video workflow as the next donor for future low-VRAM optimization. - Keep `optimized/image-to-video/minimax-h3-extender-3070-r2v-480p-clip-by-clip/` as the first `MiniMax H3 Extender` path to compare against the compact `Contex Loop` masters. - Use `optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/` as the best-current default whenever Chris wants a `MiniMax H3` `reference-to-video` continuation workflow. +- Use the new same-day `javawock7618` `INT8 R2V` import as the main donor for future `best-current-r2v` revisions if the current master needs another pass. ## Ref2VA prompt guidance @@ -44,3 +51,4 @@ - whether the Turbo LoRA workflow is just technically possible or actually usable - whether the source-recommended `8`-step path holds up in quality - whether MiniMax feels like a real challenger to `LTX` locally +- whether the same-day `INT8 R2V` donor stays practical on the `3070` once the helper stack is trimmed to the minimum diff --git a/optimized/notes/reviews/2026-08-20-daily-review.md b/optimized/notes/reviews/2026-08-20-daily-review.md new file mode 100644 index 0000000..df681e0 --- /dev/null +++ b/optimized/notes/reviews/2026-08-20-daily-review.md @@ -0,0 +1,27 @@ +# 2026-08-20 Daily Review + +## Scope + +Reviewed the current MiniMax, Z-Image, LTX, and FLUX.2-klein source packs against the maintained optimized layer. + +## Repo updates + +- Updated `optimized/image-to-video/README.md` so the category index now lists the maintained `MiniMax H3` best-current R2V master. +- Refreshed `optimized/image-to-video/minimax-h3-best-current-r2v-master-480p-3x10s/README.md` with the current Ref2VA prompt hierarchy and the same-day MiniMax donor context. + +## Notes updated + +- `optimized/notes/minimax.md` +- `optimized/notes/ltx.md` +- `optimized/notes/zimage.md` +- `optimized/notes/flux2klein.md` + +## Best first test + +- Validate the maintained `MiniMax H3` best-current R2V master on the `3070` with the new prompt hierarchy, then compare it against the same-day `javawock7618` `INT8 R2V` donor if time allows. + +## What still needs proof + +- Whether the same-day `MiniMax H3` donor is actually practical on Chris's box. +- Whether `Z-Image` edit workflows meaningfully beat the current FLUX edit baseline. +- Whether `LTX-2.5` reference control is strong enough to justify a future promoted master. diff --git a/optimized/notes/zimage.md b/optimized/notes/zimage.md index 6eecc99..0783fad 100644 --- a/optimized/notes/zimage.md +++ b/optimized/notes/zimage.md @@ -6,19 +6,29 @@ ## What to look for -- strong `text-to-image` workflows +- strong official `text-to-image` workflows - reference and edit workflows - outpaint / inpaint workflows - high-quality upscaling or restoration-adjacent workflows - any workflow that clearly beats a comparable `FLUX` path in practical image quality or simplicity +## Strong current source workflows + +- `models/zimage/z-image-turbo-official-example/` +- `models/zimage/z-image-turbo-inpainting-community/` +- `models/zimage/z-image-turbo-controlnet-union-community/` +- `models/zimage/qujincheng-z-image-lora-facetailer/` for face-detail follow-up, not as a core generator + ## Current heuristics - Prefer real workflow assets over showcase pages. - Favor workflows with clear node requirements and portable imports. - Treat `Z-Image` as especially relevant for image-generation and image-edit categories, not video categories. +- Keep the official example as the clean text-to-image baseline, then compare the community inpaint and controlnet-union graphs against it for real edit gains. +- Treat facetailer-style graphs as post-pass helpers unless they prove they can stand alone. ## What still needs proof - which `Z-Image` workflows are genuinely stronger than the best current `FLUX` image-generation paths - whether `Z-Image` should become a category-master candidate for `text-to-image` or `image-edit` +- whether the inpaint or union-style community graph is the better maintained edit baseline on Chris's machine