| name | sienna-nomad-i2v-anchor-generator |
| description | Generates I2V anchor image prompts from existing video prompts for the Sienna Nomad project. Each video prompt requires a corresponding static first-frame image to be rendered before LTX 2.3 image-to-video generation can begin. |
Sienna Nomad — I2V Anchor Image Prompt Generator
Purpose
LTX 2.3 is image-to-video. Every Prompt_NN_Video_XX.md file needs a corresponding anchor image — the static first frame that LTX uses as its i2v seed. This skill generates those anchor prompts and their JSON payloads.
Render order is mandatory:
- ALL
_ANCHOR.json payloads → rendered first via z-image
- Resulting images → stored as
Prompt_NN_Video_XX_ANCHOR.png
- Only THEN: LTX 2.3 video generation with each anchor image as i2v input
Naming Convention
| File | Purpose |
|---|
Prompt_NN_Video_XX.md | LTX video motion prompt (source) |
Prompt_NN_Video_XX_ANCHOR.md | Static first-frame image prompt (derived) |
JSON_PAYLOADS/Prompt_NN_Video_XX_ANCHOR.json | z-image payload for the anchor |
Prompt_NN_Video_XX_ANCHOR.png | Rendered anchor image (output) |
The _ANCHOR suffix is the only difference in name. This makes the pairing unambiguous.
Trigger Conditions
- When asked to "create anchor prompts for [episode]"
- When asked to "set up i2v for [episode]"
- When new video prompts are added to any episode folder
- Before any LTX 2.3 video generation run for any episode that lacks anchor images
Step-by-Step Workflow
Step 1 — Find All Video Prompts
import os, glob
episode_dir = "/Users/zgbot/Desktop/Sienna_Nomad_Project/04_Prompt_Library/EP{NN}_{Name}/"
video_prompts = sorted(glob.glob(os.path.join(episode_dir, "Prompt_*_Video_*.md")))
Skip any file that already has a corresponding _ANCHOR.md to avoid overwriting.
Step 2 — Extract First-Frame Information from Video Prompt
Read the video prompt and extract these fields for the anchor:
From the video prompt, take:
- EXPOSURE time window (same)
- CORE SUBJECT (identical — same character, same scene)
- LIGHTING/ENVIRONMENT (identical — same lighting state)
- ATMOSPHERE (identical)
- TEXTURE/MATERIALITY (identical)
- OPTICAL/CAMERAS (take lens and format, remove motion-specific notes)
Convert FLUX_DIR_COMMAND:
- REMOVE: camera_motion (there is no motion in a still frame)
- KEEP: lighting tags
- KEEP: lens tags
- KEEP: texture tags
- ADD: composition tags describing the STARTING FRAME position
For camera position in the anchor:
- The anchor shows the OPENING POSITION of the camera before any movement begins
- Example: if video says "slow lateral pan starting from the left ridge" →
anchor shows the left ridge, camera stationary, before the pan starts
- If video has a drone shot descending → anchor shows the HIGH starting altitude
- If video has a push-in → anchor shows the wide starting distance
Step 3 — Write the Anchor Prompt File
Format for Prompt_NN_Video_XX_ANCHOR.md:
# Prompt_NN_Video_XX_ANCHOR
[TYPE: I2V_ANCHOR | EXPOSURE: {same as video} | LINKED_VIDEO: Prompt_NN_Video_XX.md]
**ANCHOR ROLE:** Static first-frame image. This image is rendered by z-image and used as the i2v seed for `Prompt_NN_Video_XX.md`. Must be rendered BEFORE the video generation run.
**CINEMATIC CONTEXT:**
{Describe the opening frame — what the camera sees at the exact moment before any motion begins. Write as a still photograph, not a motion description.}
**VISUAL BLUEPRINT:**
- **CORE SUBJECT:** {Identical to video prompt — full Sienna + Aura description}
- **LIGHTING/ENVIRONMENT:** {Identical to video prompt}
- **ATMOSPHERE:** {Identical to video prompt}
- **TEXTURE/MATERIALITY:** {Identical to video prompt}
- **OPTICAL/CAMERAS:** {Lens, format, aperture — same as video. Remove motion notes.}
- **COMPOSITION:** {Opening frame composition — rule of thirds, subject placement, framing}
**FLUX_DIR_COMMAND:**
[composition: {opening_frame_composition}]
[lighting: {same lighting tags from video}]
[lens: {same lens tags from video}]
[texture: {same texture tags from video}]
Key rules for the anchor prompt:
- NO
camera_motion tag — it is a still image
- NO motion language in CINEMATIC CONTEXT ("the camera drifts", "sweeps", "pans" — remove these)
- The COMPOSITION tag must describe the exact opening frame position
- Character anchors (Sienna + Aura descriptions) must be identical to the source video prompt
- The EXPOSURE time window must match exactly — this is how LTX knows the lighting state
Step 4 — Determine Resolution from the Video Prompt
The anchor image MUST match the video output resolution. LTX 2.3 conditions on the anchor — a resolution mismatch causes stretching/cropping on the first frame.
Derive resolution from the CINEMATIC CONTEXT and OPTICAL/CAMERAS fields:
| Signal in video prompt | Resolution | Width × Height |
|---|
| Drone / wide / establishing / landscape / 24mm–35mm | Landscape 16:9 | 1024 × 576 |
| Portrait / close-up / intimate / vertical / 85mm+ | Portrait 9:16 | 576 × 1024 |
| Unclear | Default landscape | 1024 × 576 |
Record the chosen resolution in the _ANCHOR.md file under a RENDER_SPEC line:
[RENDER_SPEC: 1024×576 | MODEL: z-image | LINKED_VIDEO: Prompt_NN_Video_XX.md]
Step 5 — Generate the JSON Payload
The anchor JSON uses hermes_z_image_api.json from ~/workflows/. Read it via MCP read_workflow, then inject.
Inject the anchor prompt text by tracing from KSampler positive input to find the CLIPTextEncode node ID, then set payload["prompt"][pos_node_id]["inputs"]["text"].
Set resolution by updating the EmptySD3LatentImage node's widgets_values to the dimensions derived in Step 4:
- Landscape:
[1024, 576, 1]
- Portrait:
[576, 1024, 1]
Set filename_prefix in the SaveImage node's widgets_values to:
Prompt_NN_Video_XX_ANCHOR
Save as: JSON_PAYLOADS/Prompt_NN_Video_XX_ANCHOR.json
Step 5 — Confirm Output and Report
After generating all anchors for an episode, report:
EP{NN}_{Name} — Anchor Generation Complete
Created {N} anchor pairs:
Prompt_{NN}_Video_01 → Prompt_{NN}_Video_01_ANCHOR.md + .json ✓
Prompt_{NN}_Video_02 → Prompt_{NN}_Video_02_ANCHOR.md + .json ✓
Prompt_{NN}_Video_03 → Prompt_{NN}_Video_03_ANCHOR.md + .json ✓
RENDER ORDER FOR THIS EPISODE:
PHASE 1 (image generation): Submit all _ANCHOR.json payloads to z-image
PHASE 2 (video generation): After all anchor PNGs are confirmed rendered,
submit LTX video prompts with anchor images as i2v input
Anchor PNG expected output location: ComfyUI output folder → named Prompt_{NN}_Video_XX_ANCHOR_XXXXX.png
Pitfalls
Do NOT create anchor prompts for Photo prompts — only _Video_XX.md files get anchors. Photo prompts are standalone.
Do NOT use the video prompt text directly as the anchor text. The camera_motion and motion language must be removed. A video prompt fed into z-image will produce a blurry, motion-artifact image.
Maintain character anchor consistency. The Sienna + Aura description in the anchor must be word-for-word identical to the source video prompt. Any drift here creates visual inconsistency when LTX animates from the anchor.
The EXPOSURE time must match exactly. LTX conditions on lighting state from the anchor image. If anchor says 04:30 and video says 05:00, the lighting will fight the anchor.
Anchor resolution must match video dimensions. Use 1024×576 for landscape shots, 576×1024 for portrait. Never use square (1024×1024) — LTX will stretch it. Derive from the video prompt's CINEMATIC CONTEXT (see Step 4 table).
Gold Standard Example
⚠️ Every anchor file MUST look like this. No N/A values. No placeholder text. Every field fully written.
The completed gold standard anchor is at:
/Users/zgbot/Desktop/Sienna_Nomad_Project/04_Prompt_Library/EP01_Dawn_Ritual/Prompt_01_Video_01_ANCHOR.md
Read that file before generating any anchor. Use it as the template for every other anchor you write.
What the gold standard does:
- Pulls ALL detail from the source video prompt — lighting, atmosphere, texture, lens, character description
- Writes a full paragraph CINEMATIC CONTEXT describing the frozen opening frame
- Every VISUAL BLUEPRINT field is fully written — never N/A, never placeholder
- FLUX_DIR_COMMAND has composition tags describing the static frame position (not motion)
- Character description matches source exactly: "Sienna (30s, sun-kissed skin, freckles, Golden-Rust hair with blonde highlights) and Aura (Sleek, muscular Vizsla with a Golden-Rust coat)"
What killed the previous batch:
Every anchor was written with N/A for all fields. z-image had nothing to work with and generated generic studio shots. The model cannot invent content — you must provide it everything.
