Experiment ID: EXP-TLC-001
Status: Completed exploratory pass; character sheets and style remain candidate
Models: Anima Base v1.0 stills; Wan 2.2 14B I2V with LightX2V four-step LoRAs
Review system: SmartGallery / Media Depot

Questions

  1. Can two recurring children remain recognizable across unrelated environments using text prompts alone?
  2. Which visible anchors survive independent Anima generations?
  3. Which prompt patterns produce clean, full-bleed editorial frames?
  4. Can the visual family of the Mia Reading Nook image be repeated through a fixed text style block?
  5. Can the local dual-A4000 ComfyUI system execute two Wan 2.2 14B Turbo I2V tests in parallel?

Still-image configuration

Field Value
Runtime Local dual-instance ComfyUI
Model anima-base-v1.0.safetensors
Text encoder qwen_3_06b_base.safetensors
VAE qwen_image_vae.safetensors
Dimensions 864 × 1152, 3:4 portrait
Steps 30
CFG 4
Sampler / scheduler er_sde / simple
Batch strategy Independent seed per frame; one queue per GPU

Character direction tested

The first pass used the working names Lumi and Pip and deliberately exaggerated their contrast: twin navy buns and teal eyes for Lumi; a large auburn cowlick, scarf, and satchel for Pip. The silhouettes read clearly, but the designs felt more like stylized mascots than ordinary recurring children. Lumi also drifted more than Pip across facial structure, skin tone, clothing, pendant size, and body proportions.

The owner approved the names Mia and Henry and requested familiar, everyday-child styling.

Mia candidate

  • Chestnut-brown chin-length bob with a simple side part.
  • One small teal hair clip.
  • Warm brown eyes and light freckles.
  • Teal zip jacket, cream T-shirt, denim pants, cream sneakers.
  • Tiny round pendant as a secondary detail.
  • Calm, observant movement language.

Henry candidate

  • Compact tousled medium-brown hair with one small cowlick.
  • Warm brown eyes and light freckles.
  • Mustard zip jacket, navy T-shirt, denim pants, simple sneakers.
  • Short burnt-orange neckerchief and small blue satchel as secondary details.
  • Energetic, physical movement language.

Do not prompt a height relationship between Mia and Henry. The experiment showed that explicit height language can exaggerate scale differences and make one child read as substantially older.

Candidate reference images

Purpose Candidate
Mia solo Mia-ReadingNook-v007_00001_.png
Henry solo Henry-Soccer-v007_00001_.png
Henry alternate Henry-Pond-v007_00001_.png
Pair Mia-Henry-AutumnWalk-v007_00001_.png
Pair alternate Mia-Henry-Farm-v008_00001_.png

These are not locked reference sheets. They are the strongest candidates from this pass for the next reference-guided experiment.

Anima findings

What worked

  • Strong, simple anchors survive better than personality adjectives.
  • Mia's bob, teal clip, teal jacket, brown eyes, and freckles formed a much more stable everyday design than the twin-bun concept.
  • Henry's brown cowlick, mustard/navy block, and orange accent remained broadly recognizable.
  • The teal-led Mia and mustard-led Henry remained separable in shared scenes.
  • Ordinary, lived-in locations produced stronger emotional readability than abstract creation imagery.
  • Two GPUs processed independent still candidates reliably in parallel.

Failure patterns

  • Solid-color borders, white margins, circular vignettes, floating dioramas, and poster-like negative space.
  • Two-panel or duplicated scenes when a prompt implied multiple sequential actions.
  • Clothing, pendant, scarf, satchel, skin tone, eye shape, and body-proportion drift under text-only generation.
  • Complex hand and prop actions increased anatomy defects.
  • Text occasionally appeared in maps, exhibit labels, or artwork despite a no-text instruction.

Prompt rules adopted

  1. Begin with single continuous camera image, one frame, one moment.
  2. Require a full-bleed environment from corner to corner.
  3. Use one simple, visible physical action.
  4. Name exactly which characters are visible.
  5. Put stable hair, face, and wardrobe blocks before environmental detail.
  6. Ban borders, margins, vignettes, dioramas, posters, split screens, panels, storyboards, duplicated scenes, text, and watermarks in the model adapter.
  7. Do not state a height relationship.
  8. Treat pendant, scarf, and satchel continuity as secondary until references are conditioned directly.

Style test

Mia-ReadingNook-v007_00001_.png became the working style target. Ten new images reused one verbatim style block across different characters, environments, lighting, and actions.

Closest text-only matches:

  • Style-Mia-Breakfast-v009_00001_.png
  • Style-Mia-FlowerShop-v009_00001_.png
  • Style-Henry-Attic-v009_00001_.png
  • Style-Pair-Cookies-v009_00001_.png
  • Style-Pair-CabinMap-v009_00001_.png
  • Style-Pair-Porch-v010_00001_.png

Observed drift:

  • One frame moved toward glossy 3D rendering.
  • Cool or rainy environments changed the palette and line treatment.
  • Some frames became flatter or more graphic despite identical style text.

Conclusion: fixed text can create a recognizable style family, but it does not lock the exact Reading Nook rendering. The next style experiment should condition on an approved image reference or test a dedicated style LoRA.

Prebuilt Anima LoRA tests

Three prebuilt adapters were installed in the shared ComfyUI model tree and tested with Anima Base:

Adapter Type Test result
anima-highres-aesthetic-boost.safetensors Official aesthetic enhancer Improved individual frames but shifted between glossy 3D, painterly rendering, and outlined illustration
anima-greg-rutkowski-style.safetensors Official painterly style Polished and attractive, but the children often read older
noobai-style-anima-lora-final.safetensors Community anime style Most consistent rendering family across characters, actions, lighting, and environments

Controlled comparison

The v012 comparison rendered Mia and Henry as baseline plus all three adapters. The v013 comparison then held source character text, LoRA strength, dimensions, scene seed, and sampling settings constant while testing Aesthetic Boost against NoobAI in four new environments.

Common settings:

Field Value
Model anima-base-v1.0.safetensors
Dimensions 1024 × 1024
Steps / CFG 30 / 4
Sampler / scheduler er_sde / simple
LoRA strength 0.75
NoobAI trigger noobai-style

Result: Aesthetic Boost is an enhancer rather than a dependable style lock. NoobAI maintained the most coherent look across warm and cool scenes, interior and exterior environments, and Mia and Henry. It became the current prebuilt style-adapter candidate, not an approved final style.

Ten-frame NoobAI consistency batch

The v014 batch rendered four Mia solos, four Henry solos, and two shared scenes with unique seeds and varied camera distances, actions, props, lighting, and environments.

Output folder:

LanternClub/NoobAIConsistency

Observed:

  • The clean anime/storybook rendering remained strong across all ten frames.
  • Henry was more stable than Mia in face, hair, clothing, and apparent age.
  • Mia's main identity drift was hairstyle: bangs and side-part details varied.
  • Pair scenes preserved both children, but requested shared actions could be ignored or simplified.
  • Minor failures included malformed music mallets, an extra hair clip, pseudo-writing, and a broad plain band at the top of one art-studio frame.
  • A style LoRA improves the rendering family; it does not replace reference-conditioned character identity.

Wan 2.2 14B Turbo I2V tests

Two full-frame stills were uploaded to separate local ComfyUI instances and animated concurrently.

Test Source Motion intent Prompt ID Result
Snow Village Pair-SnowLanterns-v005-FullFrame_00001_.png Slow forward walk, scarf movement, lantern sway and flicker, falling snow, gentle backward track f5036d1f-362f-4bf8-86fc-dc94b5a8098c Technical success
Pip Bakery Pip-Bakery-v005-FullFrame_00001_.png One restrained kneading cycle, blink, cloth settling, oven flicker, slight push-in c178465d-fc56-4c0f-b6a4-ac7c916be7d4 Technical success

I2V configuration

Field Value
Models Wan 2.2 I2V 14B FP8 high-noise and low-noise
Acceleration wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors and low-noise counterpart
Steps / split 4 / 2
CFG 1
Sampler / scheduler Euler / simple
Output 480 × 640, 5 seconds, 16 fps
Runtime Approximately 2 minutes 20 seconds per test while running concurrently

The first tests prove that the workflow executes successfully on both A4000s. They do not yet approve motion quality, identity preservation, or Turbo as the final production setting.

Six-second square no-prompt batch

All ten v014 NoobAI consistency frames were animated as v015 with no text conditioning.

Field Value
Source The ten v014 NoobAI PNGs
Dimensions 512 × 512
Frames / fps 97 / 16
Effective container duration Approximately 6.063 seconds
Positive prompt Empty
Negative prompt Empty
CFG 1
Steps / split 4 / 2
Output folder LanternClub/WanI2VNoPrompt6s
Seed range 26072715012607271510

Result: unprompted Wan motion looked attractive and natural overall, and faces, clothing, and style remained reasonably stable. However, the children repeatedly moved their lips as though speaking. Wan also improvised or morphed props, including the apple, paintbrush, and paper pinwheel.

Short no-talking prompt test

Six sources were rerun as matched v016 comparisons using the same first frames, noise seeds, dimensions, duration, and Turbo workflow. The only conditioning change was this exact positive prompt:

Silent natural motion. Preserve the exact mouth shape; no talking or lip-sync.

The negative prompt remained empty. Output folder:

LanternClub/WanI2VNoTalking6s

Prompt IDs:

  • Mia Art Studio: dacc8e86-5111-464e-a602-f4359e815b52
  • Henry Music Room: c6923035-e37d-48b4-bd67-cba6743bd7ae
  • Mia Museum Butterfly: 33134484-8221-42e0-9ec1-92a712c80740
  • Henry Apple Orchard: 9605ed8f-4a93-42b7-8947-dca6df2c882d
  • Pair Market: 14be613c-5923-4af3-b7a0-22cccae68f8d
  • Pair Cabin Porch: e5d62451-d258-4e00-adb9-cd2aef9a1c35

Result: lip movement persisted. At the current Turbo settings, that short instruction did not override Wan's tendency to infer speech-like facial motion. Do not treat “no talking” as a solved prompt control.

Decisions and next experiments

Approved

  • Character names: Mia and Henry.
  • Do not prompt a height relationship.
  • Reject solid edges, blank margins, vignettes, dioramas, split panels, and duplicated scenes.

Candidate

  • The revised everyday Mia and Henry appearances.
  • Mia Reading Nook as the style target.
  • NoobAI at strength 0.75 as the strongest prebuilt style-consistency adapter tested so far.
  • The current LightX2V four-step Wan I2V workflow.

Next

  1. Select exact Mia, Henry, and pair references.
  2. Build the required turnaround, expression, action, seated, and neutral two-shot pack.
  3. Test reference-conditioned character generation.
  4. Test closed-mouth first frames with blank Wan prompts against the existing open-mouth sources and reused noise seeds.
  5. Test stronger concise mouth-restraint wording at CFG 1.
  6. Test the same source and seed at CFG 1.5–2.0, watching for Turbo quality degradation.
  7. Compare the same approved first frame in Wan Turbo and the full 20-step workflow.
  8. Review face, hands, accessory continuity, texture crawl, flicker, prop morphing, and mouth motion before selecting an I2V workflow.