qwen samples

h3-center/docs/MINIMAX_AUDIO_SYNC_AND_C2V_MUSIC_VIDEOS.md

@@ -104,3 +104,13 @@ If you take the full uncut 60s MP3 track and forcibly overlay it across the join
 2. Across 4 chained C2V clips, the aggregate video duration may differ from the raw continuous MP3 timeline by fractions of a second (e.g. 59.71s vs 59.92s).
 3. Even a ~100–200ms discrepancy results in noticeable lip-sync drift by Clip 3 and Clip 4.
 4. **Best Practice:** Always use the native C2V audio output (`MiniMax_H3_Master.mp4`), as it is sample-locked to the generated mouth movements.
+
+---
+
+## 6. Silent Reaction & Coverage Shots: Node Disconnection Protocol
+
+When generating non-singing, silent reaction shots, listening coverage, or ambient cutaways to intercut into a scene:
+- **Never leave the `LoadAudio` / `ref_audios` node connected** while instructing the character to be silent in the text prompt.
+- Active audio latent conditioning causes MiniMax-H3's cross-attention to automatically animate mouth movements and phoneme articulation, overriding negative text instructions like `"No dialogue"` or `"Lips sealed"`.
+- **The Rule:** Physically disconnect the audio reference node for all listening, reaction, or breathing takes. Only wire the audio input when the on-screen character is actively articulating vocals.
+

h3-center/docs/MINIMAX_H3_HYBRID_LORA_REFMOD_CONTAINERS_GUIDE.md

@@ -364,3 +364,4 @@ Using the new `MiniMaxH3UnifiedLoraRefModLoader`, one single node loads both the
 - [MiniMax-H3 RefMods Installation & Quickstart Guide](MINIMAX_H3_REFMODS_INSTALLATION_AND_USAGE_GUIDE.md)
 - [MiniMax-H3 Prompting Guide](MINIMAX_H3_PROMPTING_GUIDE.md)
 - [FLUX.2 Klein9 RefMod Architecture Guide](../../klein9/docs/FLUX2_KLEIN9_REFMODS_GUIDE.md)
+- [Jinx identity LoRA epoch sweep & RefMod stack](lora/README.md) — running LoRA weights and RefMod latents as two files; combined stack wins

h3-center/docs/MINIMAX_H3_PROMPTING_GUIDE.md

@@ -650,6 +650,52 @@ Treat MiniMax-H3 not as a one-click scene generator, but as a **tool for produci
 2. **Simplified Prompts**: Prompts work best with **one clear action, clear character position, clean reference frame, clear camera command, and focused dialogue**. Change one variable at a time when troubleshooting.
 3. **Editing Hides Inconsistencies**: Cut between complementary angles (Shot -> Reaction -> Reverse -> Medium). Carrying dialogue across cuts (J/L cuts) and cutting before a generation starts to drift or degrade makes the overall film feel seamless and professionally directed.
 
+### Rule 56: Global Resolution & Pixel Geometry Lock Across Timeline Cuts (September 2026)
+**The Problem:**
+Mixing generation resolutions (e.g. 1.0 MP vs 2.0 MP, or 1024x768 vs 1440x1080) across different shots in the same scene—even when they share the exact same nominal aspect ratio (e.g. 4:3 or 16:9)—creates subtle discrepancies in active image area, tensor patch padding, and scaling behavior.
+- **The Artifact:** When clips are brought into an NLE (Premiere Pro, DaVinci Resolve) or processed through upscalers, pillarbox/letterbox black borders subtly shift thickness, jump slightly between left and right edges, or cause micro-scale jumps on cuts.
+- **The Cause:** Aspect ratio is the mathematical proportion (shape), while resolution is the exact physical pixel geometry. DiT spatio-temporal VAEs pad tensors to patch divisibility (multiples of 16/32). Different pixel buckets generate slightly different padding envelopes.
+- **The Solution / Rule:**
+  1. **Lock Exact Resolution Globally:** Choose one fixed pixel resolution (e.g., exactly `1024x768` for 4:3 TNG retro framing, or `1184x672` / `1344x768` for 16:9 widescreen) for **every single shot** across the entire scene/project from day one.
+  2. **VRAM-Calibrated Baseline:** If longer shots or complex multi-subject scenes risk VRAM Out-Of-Memory (OOM) on a 24GB GPU (RTX 4090/3090), set the global standard to a stable resolution (e.g. 1.0 MP) for all shots rather than generating some at 2.0 MP and others at 1.0 MP.
+
+### Rule 57: Physical Disconnection of Voice/Audio Reference Nodes for Silent & Reaction Shots (September 2026)
+**The Problem:**
+Characters in silent reaction shots or listening coverage frequently mutter, flap their lips, or vocalize alien gibberish, even when the text prompt contains extensive silence mandates (`"No dialogue"`, `"Lips remain sealed"`, `"Jaw remains still"`).
+- **The Mechanical Root Cause:** If an audio file or voice-reference node is connected to the graph's `ref_audios` or audio conditioning input, MiniMax-H3's cross-attention mechanisms detect an active audio stream and prioritize animating mouth/lip movements to match audio latents, overriding text-based negative instructions.
+- **The Solution — Hardware/Graph-Level Node Disconnection:**
+  1. **Disconnect on Silent Takes:** For any shot where a character is listening, thinking, reacting, or silent, **physically disconnect the audio/voice reference node in ComfyUI**.
+  2. **Connect Only on Active Dialogue:** Only route the audio/voice reference into `MiniMaxH3ReferenceToVideo` when the on-screen character in that specific shot is actively vocalizing dialogue.
+  3. **Result:** Silent reaction shots become 100% stable with naturally sealed resting facial posture and zero phantom mouth flutter.
+
+### Rule 58: Performance Calibration for "Quiet Acting" vs "Big Acting" (September 2026)
+**The Problem:**
+High-energy scenes (ship alarms, shouting, running, explosions) give generative models obvious, large-scale motion vectors. Conversely, **"quiet acting"** (intimate drama, subtle remembrance, restrained dialogue, repressed emotion, micro-expressions, pensive eye gaze) is far more fragile and challenging.
+- **Common Failure Modes in Subtle Drama:**
+  - A subtle warm micro-smile turning into an exaggerated, theatrical grin.
+  - A thoughtful pause dragging half a second too long into an uncanny mannequin freeze.
+  - Excessive, jittery eye saccades or unnatural head bobbing.
+  - Restrained lines being delivered with melodrama or soap-opera cadence.
+- **The Directing Mindset — Multiple Takes & Curatorial Discipline:**
+  1. **Treat H3 as an Actor, Not a One-Shot Render:** Direct the model across multiple takes. The difference between an Oscar-worthy subtle performance and an unusable take can be a micro-second pause or 2mm of lip curvature.
+  2. **Expect a Larger Reject Pile for Quiet Drama:** Do not assume a technically glitch-free generation is "the shot." Regenerate specifically for emotional authenticity and performance restraint.
+  3. **Prompt Specificity Over Novel Length:** Keep prompts brief, grounded, and focused on physical posture and subtle eye intent rather than abstract emotional essays.
+
+### Rule 59: Coverage-Driven Television Production Workflow & Continuity Rescue (September 2026)
+**The Shift: From "Generating a Scene" to "Directing TV Coverage for the Edit"**
+Move away from attempting to generate a 30–60 second scene in a single monolithic prompt. Approach production like a traditional television director capturing modular coverage:
+1. **The 4-Shot Coverage Package:**
+   - **A-Roll (Speaker MCU):** Medium close-up of Character A delivering line (Voice reference connected).
+   - **B-Roll (Listener Reaction):** Medium close-up of Character B reacting/processing (Voice reference disconnected).
+   - **C-Roll (Inserts & Cutaways):** Contextual props or environmental elements (e.g. Picard holding the Ressikan flute, tea cup on desk, PADD display, engine console).
+   - **D-Roll (Pensive Beat / Breath):** 2–3 seconds of silent character reflection before or after a critical beat.
+2. **Editorial Assembly (J-Cuts & L-Cuts in Premiere / Resolve):**
+   - Lay down audio dialogue continuously across cuts.
+   - Let Character A's line finish over Character B's silent reaction shot (**L-Cut**).
+   - Cut to Character A listening before Character B begins speaking off-screen (**J-Cut**).
+3. **Continuity Masking:**
+   - Reaction shots and insert cutaways naturally break up generational variances in background lighting, micro-clothing details, posture shifts, or slight face likeness drift between different generation batches. The edit does the heavy lifting of continuity.
+
 ---
 
 ## 3. Multi-Scene Movie Production Workflow (`docs/movies/movieXYZ/`)

h3-center/docs/MINIMAX_H3_REFMODS_INSTALLATION_AND_USAGE_GUIDE.md

@@ -116,3 +116,6 @@ For multi-character scenes, assign `mod_1` and `mod_2` in the loader and define
 ## 7. Advanced: Unified Hybrid LoRA + RefMod Containers
 To package and distribute both fine-tuned LoRA weights and pre-encoded RefMod latents in a single `.safetensors` file compatible with both standard ComfyUI LoRA loaders and specialized dual loaders, see:
 👉 **[`MINIMAX_H3_HYBRID_LORA_REFMOD_CONTAINERS_GUIDE.md`](MINIMAX_H3_HYBRID_LORA_REFMOD_CONTAINERS_GUIDE.md)**
+
+Until a hybrid file ships, load identity LoRA and RefMod as two adapters. Empirically, **RefMod + identity LoRA beats either alone**; Jinx epoch 60 is enough (no gain at 80 / final). Clips and write-up:
+👉 **[`lora/README.md`](lora/README.md)**

h3-center/docs/MINIMAX_H3_REFMOD_STACKING_AND_MULTISUBJECT_GUIDE.md

@@ -17,6 +17,7 @@ This guide provides the complete architectural, mathematical, and practical expl
   - [Reference Audio Lip-Syncing & Multi-Clip C2V Music Videos](MINIMAX_AUDIO_SYNC_AND_C2V_MUSIC_VIDEOS.md)
   - [MiniMax-H3 Video Prompt Writing Guide](MINIMAX_H3_PROMPTING_GUIDE.md)
   - [C2V Continuous Generation & Long-Shot Chains Guide](C2V_CONTINUOUS_CHAIN_GUIDE.md)
+  - [Jinx identity LoRA epoch sweep & RefMod stack](lora/README.md) — empirical A/B: epoch 60 enough; RefMod + LoRA best; clips in [`lora/`](lora/)
 
 ---
 

h3-center/docs/MINIMAX_REFMODS_VS_REFERENCE_IMAGES.md

@@ -131,3 +131,4 @@ This gives the ultimate combination: **RefMods ensure flawless character facial
 3. **Record locked RefMods in persona look sheets** (`docs/personas/<name>.md`) and character profiles (`docs/characters/<name>.md`).
 4. **Avoid Generic Beauty Buzzwords in Prompts:** Do NOT use generic beauty buzzwords like `"attractive Caucasian young woman"`, `"soft oval face"`, `"straight slender nose"`. These keywords activate MiniMax's generalized beauty prior and overwrite unique RefMod facial features. Always prompt with authentic biometric traits directly (see [`MINIMAX_H3_PROMPTING_GUIDE.md`](MINIMAX_H3_PROMPTING_GUIDE.md)).
 5. **Deprecate raw image lists (`referenceImages: [...]`) in favor of RefMod loader nodes** for all new production scripts and workflows.
+6. **When an identity LoRA exists, stack it with the RefMod.** Jinx terrace A/B (same prompt/seed): RefMod-only, LoRA-only, and combined — combined is the tightest likeness. Epoch 60 of that LoRA matches 80 and final. See [`lora/README.md`](lora/README.md).

h3-center/docs/lora/README.md

@@ -0,0 +1,106 @@
+# Jinx Identity LoRA Epoch Sweep & RefMod Stack
+
+_Date: 19 September 2026_  
+_Character: Jinx (Arcane) / Ella Purnell likeness_  
+_Clips in this folder are the source of truth for the conclusions below._
+
+This is a small A/B of a new **identity LoRA** (`minimaxh3_jinx_v1`) against a **single identity RefMod** (`minimaxh3_jinx_v1_refmod`), then both together. The headline result: **RefMod + LoRA is the best stack**. Epoch 60 of the LoRA is enough; later checkpoints do not help.
+
+---
+
+## Conclusions
+
+1. **Epoch 60 is enough.** Terrace isolation clips at e060, e080, and `final` are effectively the same identity and style. There is no reason to train this LoRA further for MiniMax-H3 T2V.
+2. **RefMod + LoRA wins.** Same terrace prompt and seed: LoRA-only, RefMod-only, and RefMod+e060. The combined stack has the tightest Jinx lock (braids, tattoos, face).
+3. **This is a bad test of “more realistic vs more Arcane.”** The terrace prompt asked for **live-action 35mm** and “played by Ella Purnell.” It did **not** ask for Arcane 3D. Isolation LoRA still came out Arcane-3D. RefMod-only came out closer to live-action / Purnell. Combined stack stayed Arcane-3D with better likeness. We therefore **cannot** score whether extra epochs made her “more realistic.” We can only score identity lock and whether the adapters agree with each other.
+4. **Style is not identity.** The LoRA appears to carry **Arcane look** as well as face. The RefMod carries **who she is** without forcing the painted 3D look as hard. Use them together when you want a locked Jinx. Do not use this sweep to decide live-action vs animation.
+
+---
+
+## Clip inventory
+
+All clips: **5.0s / 120 frames / 1344×768 / 24 fps**, Turbo 8-step, FirstBlockCache Fast 0.10, SageAttention, Euler / Simple, denoise 1.0.
+
+| File | Stack | Seed | What it shows |
+| :--- | :--- | :--- | :--- |
+| [`job_jinx_e060_terrace_5s_00001_.mp4`](job_jinx_e060_terrace_5s_00001_.mp4) | LoRA e060 @ 0.70, **no RefMod** | `2026091921` | Isolation. Already Arcane-3D despite a live-action prompt. |
+| [`job_jinx_e080_terrace_5s_00001_.mp4`](job_jinx_e080_terrace_5s_00001_.mp4) | LoRA e080 @ 0.70, **no RefMod** | `2026091921` | Same shot. No meaningful gain over e060. |
+| [`job_jinx_final_terrace_5s_00001_.mp4`](job_jinx_final_terrace_5s_00001_.mp4) | LoRA `final` @ 0.70, **no RefMod** | `2026091921` | Same shot. Matches e080. Training past 60 is wasted. |
+| [`job_jinx_refmod_only_terrace_5s_00001_.mp4`](job_jinx_refmod_only_terrace_5s_00001_.mp4) | **RefMod only** @ 1.0, no identity LoRA | `2026091921` | Same prompt/seed. More live-action / Purnell. Identity is weaker than the combined stack. |
+| [`job_jinx_e060_refmod_terrace_5s_00001_.mp4`](job_jinx_e060_refmod_terrace_5s_00001_.mp4) | **RefMod @ 1.0 + LoRA e060 @ 0.70** | `2026091921` | **Best terrace result.** Arcane-3D Jinx, tighter face, two braids, tattoos. |
+| [`job_jinx_e060_refmod_punch_picard_5s_00001_.mp4`](job_jinx_e060_refmod_punch_picard_5s_00001_.mp4) | RefMod + LoRA e060 + Combat V2 @ 0.52 | `2026091931` | Two-shot vs Picard (Patrick Stewart). Framing is right (mid-range profiles). Punch does not land. Mixed style: Jinx 3D, Picard live-action. |
+
+Shared terrace prompt (abbreviated): live-action 35mm golden-hour penthouse terrace, tank top, medium close-up into camera, spoken line *“Hi, it is me. This warm evening light feels perfect.”* Subject tag: `Jinx (played by Ella Purnell)` with blue braids and cloud tattoo — **no Arcane / 3D style lock in the shot block.**
+
+---
+
+## Technical: identity LoRA
+
+| | |
+| :--- | :--- |
+| Files | `minimaxh3_jinx_v1-000060.safetensors`, `minimaxh3_jinx_v1-000080.safetensors`, `minimaxh3_jinx_v1.safetensors` (`final`) |
+| Size | **149,164,872 bytes** (~142.3 MiB) each — same rank/shape across epochs |
+| Role | Weight patch on the MiniMax-H3 DiT (and CLIP if the loader applies `strength_clip`) |
+| ComfyUI path | `models/loras/MinimaxH3/test/` |
+| Loader | rgthree **Power Lora Loader**, slot after Turbo |
+| Tested strength | **0.70** model + CLIP |
+| Turbo (always on) | `MinimaxH3\special\minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` @ **1.0** |
+
+A LoRA is **not** a reference image. It changes how the transformer predicts motion and appearance. That is why isolation e060 already **overrides** a live-action prompt into Arcane-3D: the training distribution is in the weights. Extra epochs (80 / final) did not change that distribution in a way this terrace A/B can see.
+
+Practical rule from this sweep: **stop Jinx identity training at epoch 60** for H3 T2V. Keep the e060 file as the production checkpoint.
+
+---
+
+## Technical: Jinx RefMod
+
+| | |
+| :--- | :--- |
+| File | `minimaxh3_jinx_v1_refmod.safetensors` |
+| Size | **1,364,704 bytes** (~1.30 MiB) |
+| Role | Pre-encoded **identity latents** (VAE memory), not weight deltas |
+| ComfyUI path | `models/refmods/` |
+| Nodes | `MiniMaxH3RefModsLoader` → `MiniMaxH3RefModApply` |
+| Tested settings | `mod_1` = jinx, **strength 1.0**, **copies 1**, apply **retention 1.0**, **curve_direction constant** |
+
+RefMods sit in the packed DiT sequence next to text tokens (`[text | refmod latents | audio | video]`). They do not patch `lora_down` / `lora_up`. That is why RefMod-only can keep a more photographic read while still pulling braids, eyes, and tattoos: it is visual memory, not a full style finetune.
+
+Single-RefMod (not a triple ensemble) is the right comparison here. Jinx does not have a `jinx7` / `jinxlarge` / `jinxlatest` stack the way Aneta/Gosia do.
+
+---
+
+## Why RefMod + LoRA is the production stack
+
+| Adapter | What it does | Terrace result |
+| :--- | :--- | :--- |
+| LoRA only | Patches generation toward the training look | Strong Arcane style, identity OK |
+| RefMod only | Injects identity latents into conditioning | More live-action, identity looser |
+| **Both** | Weights pull style/motion; latents pin the face | **Best likeness, coherent Jinx** |
+
+This matches the hybrid-container idea: LoRA keys patch MODEL/CLIP; `refmod.*` keys are conditioning. Until a unified file exists, load them as two files in the graph (Power LoRA Loader + RefModsLoader). See [Hybrid LoRA + RefMod Containers](../MINIMAX_H3_HYBRID_LORA_REFMOD_CONTAINERS_GUIDE.md).
+
+Do **not** read the Picard punch as a likeness test. Combat V2 was on the stack; the hit never connected. Use it only as a mixed-style two-shot example.
+
+---
+
+## What a fair style A/B would need
+
+To know whether the LoRA “makes her more realistic” or “more Arcane,” rerun the **same seed** with an explicit shot style, for example:
+
+- Live-action lock: `Live-action 35mm photograph, photoreal skin, Ella Purnell likeness, not 3D, not Arcane animation`
+- Arcane lock: `3D stylized Arcane / Fortiche look, painted lighting, not live-action`
+
+Until that exists, treat e060 as **identity+style** and the RefMod as **identity**. Combined is the default.
+
+---
+
+## Related docs
+
+- [RefMods installation & usage](../MINIMAX_H3_REFMODS_INSTALLATION_AND_USAGE_GUIDE.md)
+- [RefMod creation](../MINIMAX_H3_REFMOD_CREATION_GUIDE.md)
+- [RefMods vs reference images](../MINIMAX_REFMODS_VS_REFERENCE_IMAGES.md)
+- [RefMod stacking & multi-subject](../MINIMAX_H3_REFMOD_STACKING_AND_MULTISUBJECT_GUIDE.md)
+- [Hybrid LoRA + RefMod containers](../MINIMAX_H3_HYBRID_LORA_REFMOD_CONTAINERS_GUIDE.md)
+- [Prompting guide](../MINIMAX_H3_PROMPTING_GUIDE.md)
+- [Multi-stacking video suite](../multistacking/README.md)
+- Character roll (Arcane prompt, different from this terrace test): [`crossovers/docs/characters/arcane/jinx.md`](../../crossovers/docs/characters/arcane/jinx.md)

h3-center/docs/lora/job_jinx_e060_refmod_punch_picard_5s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37fa4d72cbb136b7a919109f52c8ce9a9ce01e83a97c56dbc220048ea950493d
+size 1020172

qwen/samples/edit/Qwen_GillianAnderson_DanaScully_5Ref_RoleBased_00001.png → h3-center/docs/lora/job_jinx_e060_refmod_terrace_5s_00001_.mp4

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f4d4af6986ed02a90522f33c4c727eae21794c8a2659527dabbd95354a8528df
-size 657836
+oid sha256:f7ddf9fcc90effad75db707680b616a5927ca1813dfd18cf540807c7db98e1da
+size 732456

h3-center/docs/lora/job_jinx_e060_terrace_5s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1724aee1b00da0fb3ea7750c36d0969136a98426968c146a4b56607828adc16c
+size 703865

h3-center/docs/lora/job_jinx_e080_terrace_5s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efdc29569fc66d205a6d8d64182ff2e860ee5704788337798f68d3559c882b7c
+size 724943

h3-center/docs/lora/job_jinx_final_terrace_5s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a7c5d30a71cbeb36ba2835d64997c7f45a82fab008e28585471da8f8cdef4bc
+size 699924

h3-center/docs/lora/job_jinx_refmod_only_terrace_5s_00001_.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4feb2c5e4c1e451f6f537095b34d3308f1c0afd9168b02bca83368dca23b91d2
+size 865513

h3-center/docs/lora/mieszkania.md

@@ -0,0 +1,72 @@
+Oto **zweryfikowane, aktualne i aktywne** oferty wynajmu mieszkań 3-pokojowych (od 40 m² wzwyż) położonych w wyznaczonym pasie: **od Szczepina (okolice ul. Dobrej / przedszkola) wzdłuż osi Legnickiej / Popowic aż po Dom Handlowy Astra (Gądów Mały / Kwiska / Na Ostatnim Groszu)**.
+
+---
+
+### 1. Bezpośrednio przy ul. Dobrej (Szczepin)
+
+* **[Mieszkanie 3-pokojowe – ul. Poznańska (Szczepin)](https://www.olx.pl/d/oferta/mieszkanie-na-wynajem-poznanska-3-pokoje-parter-CID3-ID1c6wfB.html)**
+  * **Lokalizacja:** ul. Poznańska (tuż obok ul. Dobrej i Pl. Strzegomskiego / „pociągu do nieba”)
+  * **Metraż:** 48 m²
+  * **Układ:** 3 pokoje (salon + 2 pokoje), kuchnia, łazienka, parter, piwnica w cenie; po generalnym remoncie, cicha okolica w głębi osiedla.
+  * **Cena:** 3 500 zł + ok. 1 000 zł (czynsz adm. + prąd wg zużycia)
+  * **Kaucja:** do uzgodnienia
+
+* **[Mieszkanie 3-pokojowe – ul. Inowrocławska (Szczepin)](https://www.otodom.pl/pl/oferta/3-pok-szczepin-winda-rozklad-dla-studenta-ID4CDfK)**
+  * **Lokalizacja:** ul. Inowrocławska (Szczepin, kilka minut pieszo od ul. Dobrej i Pl. Jana Pawła II)
+  * **Metraż:** 47 m²
+  * **Układ:** 3 pokoje, kuchnia, łazienka z toaletą, loggia, 1. piętro z windą; w pełni wyposażone (zmywarka, pralko-suszarka, płyta indukcyjna).
+  * **Cena:** 3 700 zł + 950 zł (czynsz adm. z zaliczkami na media)
+  * **Kaucja:** 5 000 zł
+
+---
+
+### 2. Odcinek środkowy: Legnicka / okolice CH Magnolia Park / Popowice
+
+* **[Nowoczesne 3 pokoje – ul. Legnicka (obok Magnolia Park)](https://www.otodom.pl/pl/oferta/na-wynajem-62-m-3-pokoje-balkon-legnicka-obok-magnolia-park-ID4D5lN)**
+  * **Lokalizacja:** ul. Legnicka (przy CH Magnolia Park, nowy budynek)
+  * **Metraż:** 62 m²
+  * **Układ:** Salon z aneksem kuchennym i wyjściem na balkon + sypialnia główna (łóżko 180 cm) + drugi pokój (gabinet/sypialnia) + łazienka (z pralko-suszarką) + przedpokój z szafą.
+  * **Cena:** 4 200 zł (do negocjacji) + ok. 1 000 zł opłat; możliwość dokupienia miejsca w garażu podziemnym.
+
+* **[Rozkładowe 3 pokoje z loggiami – ul. Białowieska 3 (Pixel House / Legnicka)](https://www.otodom.pl/pl/oferta/0prowizji-studenci-pary-rodzina-3-pokoje-rozkladowe-ul-legnicka-ID4rVRs)**
+  * **Lokalizacja:** ul. Białowieska 3 (budynek Pixel House naprzeciwko Magnolii)
+  * **Układ:** 3 pokoje (każdy z wyjściem na osobną loggię), wydzielony aneks kuchenny, łazienka z prysznicem i WC; budynek chroniony z windą.
+  * **Cena:** 4 000 zł (najem z administracją) + 350 zł zaliczki na media. **0% prowizji**.
+
+* **[3 pokoje po remoncie – ul. Białowieska 15 (Popowice)](https://www.otodom.pl/pl/oferta/3-pokoje-popowice-bialowieska-okolice-magnolii-ID4D8Ko)** *(dostępne też na [OLX](https://www.olx.pl/d/oferta/3-pokoje-popowice-bialowieska-okolice-magnolii-CID3-ID1cebgP.html))*
+  * **Lokalizacja:** ul. Białowieska 15 (Popowice, okolice Magnolii i parku)
+  * **Metraż:** 54 m²
+  * **Układ:** Salon + 2 osobne sypialnie + kuchnia + łazienka + osobne WC + balkon + piwnica. 1. piętro, parking za szlabanem.
+  * **Cena:** 2 600 zł + ok. 1 100 zł opłat (łącznie ok. 3 700 zł)
+  * **Kaucja:** 2 600 zł (bez prowizji, od osoby prywatnej)
+
+* **[3 pokoje z dużym balkonem – ul. Legnicka 118 (przy ul. Kwiskiej)](https://www.olx.pl/d/oferta/3-pokoje-z-duzym-balkonem-legnicka-cisza-parking-za-szlabanem-CID3-ID1ceAKA.html)**
+  * **Lokalizacja:** ul. Legnicka 118 (druga linia zabudowy, cicho, przy skrzyżowaniu z Kwiską)
+  * **Metraż:** 54 m²
+  * **Układ:** 3 pokoje, osobna kuchnia z AGD, łazienka z prysznicem, osobne WC, duży balkon na zieleń, parking za szlabanem dla mieszkańców.
+  * **Cena:** 2 900 zł + 713 zł opłaty do zarządcy + ok. 300 zł media
+
+* **[3-pokojowe mieszkanie – ul. Wejherowska (Popowice)](https://www.olx.pl/d/oferta/mieszkanie-3-pokojowe-54m2-do-wynajecia-wejherowska-popowice-dla-2-3-osob-CID3-ID1c2pt7.html)**
+  * **Lokalizacja:** ul. Wejherowska (blisko Parku Zachodniego i basenu Orbita)
+  * **Metraż:** 54 m²
+  * **Układ:** 3 pokoje, balkon, 1. piętro z windą, osobna kuchnia ze zmywarką, osobno łazienka i WC.
+  * **Cena:** 2 900 zł + 800 zł czynsz adm. + media (prąd ok. 170 zł, internet 70 zł)
+  * **Kaucja:** 3 000 zł
+
+---
+
+### 3. Bezpośrednio przy Domu Handlowym Astra (Gądów Mały / Na Ostatnim Groszu)
+
+* **[3 pokoje – ul. Szybowcowa (tuż przy Astrze)](https://www.otodom.pl/pl/oferta/blisko-tramwaj-balkon-i-piwnica-od-zaraz-okolice-astry-ID4CReY)**
+  * **Lokalizacja:** ul. Szybowcowa (Gądów Mały, 2 minuty pieszo do CH Astra i przystanków tramwajowych)
+  * **Metraż:** 50 m²
+  * **Układ:** 3 pokoje, balkon, piwnica, budynek z windą (wysoki parter), internet w cenie najmu.
+  * **Cena:** 3 000 zł + ok. 950 zł czynsz adm. + prąd wg zużycia
+  * **Kaucja:** 4 000 zł
+
+* **[3 pokoje rozkładowe – ul. Na Ostatnim Groszu (obok Astry)](https://www.otodom.pl/pl/oferta/na-ostatnim-groszu-rozklad-balkon-piwnica-ID4CR68)**
+  * **Lokalizacja:** ul. Na Ostatnim Groszu (przy Astrze i wylocie na Legnicką)
+  * **Metraż:** 54 m²
+  * **Układ:** 3 niezależne pokoje (salon 14,2 m² z balkonem + sypialnie 11,7 m² i 9,7 m²), osobna kuchnia z oknem i zmywarką, łazienka i osobne WC, 9. piętro z windą.
+  * **Cena:** 3 200 zł + ok. 1 000 zł opłat eksploatacyjnych
+  * **Kaucja:** 4 200 zł

h3-center/docs/multistacking/README.md

@@ -5,6 +5,9 @@ This directory contains verified `.mp4` video generation benchmarks demonstratin
 For complete architectural details, mathematical foundations, and prompting guidelines, see the main guide:
 👉 **[`../MINIMAX_H3_REFMOD_STACKING_AND_MULTISUBJECT_GUIDE.md`](../MINIMAX_H3_REFMOD_STACKING_AND_MULTISUBJECT_GUIDE.md)**
 
+Identity LoRA vs RefMod A/B (Jinx epoch 60 / 80 / final, RefMod-only, RefMod+LoRA):
+👉 **[`../lora/README.md`](../lora/README.md)** — epoch 60 is enough; **RefMod + LoRA is the best stack**.
+
 ---
 
 ## Benchmark Video Inventory