ntc-ai/anima-particle-sliders

Anima Particle Sliders

Bad Intent, Final Form, Afterimage, Moonlit, Candlelit, Dusk for Anima Turbo v1.1. Original Particle adapters and ordinary LoRA Distills, with strength stored in each file’s alpha.

Samples

Start at strength 1.0. Every current comparison is Particle 1 → Distill 1 → Off 0, with the same prompt, seed, 768 × 768 resolution, 10 Euler steps and CFG 1. These are newly rendered examples from the calibrated files linked below.

Bad Intent adds menacing expressions and close framing. Final Form adds supernatural transformations; Afterimage adds translucent, offset echoes of faces and gestures. Moonlit, Candlelit and Dusk change atmosphere. Distills are linear approximations of the original particles; their images need not match exactly.

Bad Intent

Particle · strength 1 → Distill · strength 1 → Off · strength 0

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1boy, male focus, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater and olive trousers, long sleeves, plain fabric, facing viewer, head upright, portrait, head and shoulders, eye-level camera, in a wood-paneled room, anime illustration

Seed: 29027.

Final Form

Particle · strength 1 → Distill · strength 1 → Off · strength 0

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, tan skin, medium brown skin, long straight slate-blue hair, hazel eyes, neutral expression, closed mouth, looking at viewer, wearing a tobacco-brown coat and black skirt, long sleeves, plain fabric, standing upright, facing viewer, arms at sides, upper body, waist-up view, eye-level camera, in an unadorned stone corridor with bare stone walls, anime illustration

Seed: 29001.

Afterimage

Particle · strength 1 → Distill · strength 1 → Off · strength 0

Full resolution: Particle · Distill · Off

Prompt and seed

anime illustration, one adult man age 35, light skin, short curly red hair, blue eyes, wearing a cream sweater, turning his shoulders and lifting one hand, medium close-up, soft long-exposure ghosting around the moving face and hand, one solid central person, subdued violet lighting, plain dark backdrop

Seed: 29027.

Moonlit

Particle · strength 1 → Distill · strength 1 → Off · strength 0

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, solo, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater, crew neck, facing viewer, head upright, close-up portrait, head and shoulders, face filling the frame, eye-level camera, in an unadorned stone corridor with bare stone walls, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Candlelit

Particle · strength 1 → Distill · strength 1 → Off · strength 0

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, solo, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater, crew neck, facing viewer, head upright, close-up portrait, head and shoulders, face filling the frame, eye-level camera, in an unadorned stone corridor with bare stone walls, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Dusk

Particle · strength 1 → Distill · strength 1 → Off · strength 0

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, tan skin, medium brown skin, long straight slate-blue hair, hazel eyes, neutral expression, closed mouth, looking at viewer, wearing a tobacco-brown coat and black skirt, long sleeves, plain fabric, standing upright, facing viewer, arms at sides, upper body, waist-up view, eye-level camera, in an unadorned stone corridor with bare stone walls, anime illustration, soft illumination, gentle shadows

Seed: 29001.

More freshly rendered strength-one comparisons

Bad Intent · case1

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater and olive trousers, long sleeves, plain fabric, standing upright, facing viewer, both hands on hips, cowboy shot, thighs in frame, eye-level camera, in an unadorned stone corridor with bare stone walls

Seed: 29001.

Final Form · case1

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater and olive trousers, long sleeves, plain fabric, standing upright, facing viewer, both hands on hips, cowboy shot, thighs in frame, eye-level camera, in an unadorned stone corridor with bare stone walls

Seed: 29001.

Afterimage · gesture-echo

Full resolution: Particle · Distill · Off

Prompt and seed

anime illustration, one adult man age 35, light skin, short curly red hair, blue eyes, wearing a cream sweater, turning his shoulders and lifting one hand, medium close-up, soft long-exposure ghosting around the moving face and hand, one solid central person, subdued violet lighting, plain dark backdrop

Seed: 29001.

Moonlit · case0

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, tan skin, medium brown skin, long straight slate-blue hair, hazel eyes, neutral expression, closed mouth, looking at viewer, wearing a tobacco-brown coat and black skirt, long sleeves, plain fabric, standing upright, facing viewer, arms at sides, upper body, waist-up view, eye-level camera, in an unadorned stone corridor with bare stone walls, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Moonlit · case1

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, dark skin, deep brown skin, short silver coils, brown eyes, neutral expression, closed mouth, looking at viewer, wearing a jade high-neck blouse and charcoal trousers, long sleeves, plain fabric, facing viewer, head upright, portrait, head and shoulders, eye-level camera, in a quiet courtyard, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Moonlit · case2

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, tan skin, medium brown skin, long straight slate-blue hair, hazel eyes, neutral expression, closed mouth, looking at viewer, wearing a tobacco-brown coat and black skirt, long sleeves, plain fabric, sitting upright on a plain wooden chair, facing viewer, feet flat on floor, hands resting in lap, full body, entire figure in frame, eye-level camera, in a wood-paneled room, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Moonlit · case3

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater and olive trousers, long sleeves, plain fabric, standing upright, facing viewer, both hands on hips, cowboy shot, thighs in frame, eye-level camera, in an unadorned stone corridor with bare stone walls

Seed: 29001.

Candlelit · case0

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, tan skin, medium brown skin, long straight slate-blue hair, hazel eyes, neutral expression, closed mouth, looking at viewer, wearing a tobacco-brown coat and black skirt, long sleeves, plain fabric, standing upright, facing viewer, arms at sides, upper body, waist-up view, eye-level camera, in an unadorned stone corridor with bare stone walls, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Candlelit · case1

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, dark skin, deep brown skin, short silver coils, brown eyes, neutral expression, closed mouth, looking at viewer, wearing a jade high-neck blouse and charcoal trousers, long sleeves, plain fabric, facing viewer, head upright, portrait, head and shoulders, eye-level camera, in a quiet courtyard, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Candlelit · case2

Full resolution: Particle · Distill · Off

Prompt and seed

masterpiece, best quality, safe, solo, 1girl, mature female, adult woman age 35, tan skin, medium brown skin, long straight slate-blue hair, hazel eyes, neutral expression, closed mouth, looking at viewer, wearing a tobacco-brown coat and black skirt, long sleeves, plain fabric, sitting upright on a plain wooden chair, facing viewer, feet flat on floor, hands resting in lap, full body, entire figure in frame, eye-level camera, in a wood-paneled room, anime illustration, neutral diffuse illumination, moderate shadow contrast

Seed: 29001.

Candlelit · case3

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater and olive trousers, long sleeves, plain fabric, standing upright, facing viewer, both hands on hips, cowboy shot, thighs in frame, eye-level camera, in an unadorned stone corridor with bare stone walls

Seed: 29001.

Dusk · case1

Full resolution: Particle · Distill · Off

Prompt and seed

1boy, male focus, mature male, adult man age 35, light skin, short curly red hair, blue eyes, neutral expression, closed mouth, looking at viewer, wearing a cream cable-knit sweater and olive trousers, long sleeves, plain fabric, standing upright, facing viewer, both hands on hips, cowboy shot, thighs in frame, eye-level camera, in an unadorned stone corridor with bare stone walls

Seed: 29001.

These are development examples, not a final-test benchmark. Strength 1 is the calibrated default; higher settings extrapolate and can change pose, clothing, framing or appearance.

Get the adapters

Slider	Original particles	Distilled LoRA for ComfyUI	Distilled LoRA for native runtime
Bad Intent	Download	Download	Download
Final Form	Download	Download	Download
Afterimage	Download	Download	Download
Moonlit	Download	Download	Download
Candlelit	Download	Download	Download
Dusk	Download	Download	Download

Particles: use the ComfyUI particle plugin at strength 1.0. Update the plugin for embedded-alpha files. Plugin ZIP · Setup.

Distilled LoRAs: use standard Load LoRA, MODEL strength 1.0, CLIP strength 0. No particle plugin is needed.

Slider	Particle alpha	Distill alpha	Rank
Bad Intent	12	24	8
Final Form	12	24	8
Afterimage	8	16	8
Moonlit	23.08072422	23.08072422	8
Candlelit	16	16	8
Dusk	32	96	8

Calibration is embedded in the files; do not add an external gain. Bad Intent’s old alpha-8 LoRA needs strength 3 for the same operation as the new alpha-24 file at strength 1. Lighting LoRAs carry the same alpha gain as their calibrated particle teacher; this does not claim a new fit or exact image equivalence.

Source and reproduction · Catalog · File hashes · Distillation and alpha audit.

The reusable algorithms come from the pinned shared core. Training weights and the ParticleGAN formulation are unchanged.

Historical exports and examples

The original alpha-8 exports and their samples remain available for reproduction. Their strength numbers use the original scale and do not describe the new calibrated files. The catalog’s archived_release entries retain the original paths.

Original lighting/Bad Intent particle samples · Original distill samples.

Original, uncalibrated projection-relative fitting MSE: Bad Intent 0.005434, Moonlit 0.007559, Candlelit 0.020457. These local diagnostics are not image-quality scores and do not measure the calibrated complete-denoiser edit.

How the sliders learn

These are nonlinear attention adapters for Anima Turbo v1.1. The base diffusion transformer, text encoders, text conditioner, and VAE stay frozen. Bad Intent, Candlelit and Moonlit each have their own adapter, critic, and learned particle cloud. Bad Intent uses the same game with a broader psychological-horror target: expression, pose and framing may change as well as atmosphere. Its internal training identifier remains uncanny.

The particle branch

Every self-attention and cross-attention Q, K, V, and output projection receives an additive branch: 224 projections across 28 transformer blocks. For a projection input $xx$, define

$$u=W_{\mathrm{d}\mathrm{o}\mathrm{w}\mathrm{n}}x,q=R(u),a=\mathrm{softmax}\left(\frac{q{P}^{\mathrm{\top }}}{\sqrt{4}}\right),z=aP\mathrm{.}u\; =\; W\_\{\backslash mathrm\{down\}\}x,\backslash qquad\; q=R(u),\backslash qquad\; a=\backslash operatorname\{softmax\}\backslash left(\backslash frac\{qP^\backslash top\}\{\backslash sqrt\{4\}\}\backslash right),\backslash qquad\; z=aP.$$

Here $W_{\mathrm{d}\mathrm{o}\mathrm{w}\mathrm{n}}W\_\{\backslash mathrm\{down\}\}$ maps into rank 8, $P\in {\mathbb{R}}^{128\times 4}P\backslash in\backslash mathbb\; R^\{128\backslash times4\}$ is the shared particle cloud, and $RR$ is a per-projection MLP of width 16. The branch MLP $FF$ has width 48. All hidden activations are LeakyReLU with slope 0.2; both MLPs have three hidden linear layers and a final output layer. The projection becomes

$$y=W_{0}x+s{W}_{\mathrm{u}\mathrm{p}}F([u,z])\mathrm{.}y\; =\; W\_0x\; +\; s\backslash ,W\_\{\backslash mathrm\{up\}\}F([u,z]).$$

The up matrix maps rank 8 back to the projection output width and starts at zero. The same soft routing runs during training and inference. Strength $s=0s=0$ bypasses the branch exactly. Training uses $s=1s=1$; the historical gallery included extrapolation through $s=5s=5$. Calibrated inference additionally multiplies the branch by its stored alpha/rank; the current gallery uses nominal strength 1. No training weights or losses change. For multiple sliders the deltas sum on the same input. Studio normalizes nonnegative mix proportions as $s_j=E{m}_j\mathrm{/}{\sum }_k{m}_{k}s\_j=E\; m\_j/\backslash sum\_k\; m\_k$.

An ordinary LoRA has a linear branch $BAxBAx$. The routed particle branch is nonlinear, so its full behavior cannot be merged into a fixed weight matrix. The distilled LoRAs are fitted approximations.

Paired frozen targets

For each matched neutral/positive prompt pair and seed, generate both frozen trajectories from the same initial noise. At every one of the ten sampling positions on both paths, evaluate the neutral and positive frozen velocities at that same latent state. No spatial pooling is used.

Let $v_0(z,t,c_n)v\_0(z,t,c\_n)$ and $v_0(z,t,c_p)v\_0(z,t,c\_p)$ be those frozen velocities. Let $v_{\theta }(z,t,c_n)v\_\backslash theta(z,t,c\_n)$ be the adapted neutral prediction. The desired and student edits are

$${\mathrm{\Delta }}_T=v_0(z,t,c_p)-v_0(z,t,c_n),{\mathrm{\Delta }}_S=v_{\theta }(z,t,c_n)-v_0(z,t,c_n)\mathrm{.}\backslash Delta\_T=v\_0(z,t,c\_p)-v\_0(z,t,c\_n),\backslash qquad\; \backslash Delta\_S=v\_\backslash theta(z,t,c\_n)-v\_0(z,t,c\_n).$$

The normalized residual is

$$e_t=\frac{{\mathrm{\Delta }}_S-{\mathrm{\Delta }}_T}{S_t}=\frac{v_{\theta }(z,t,c_n)-v_0(z,t,c_p)}{S_t}\mathrm{.}e\_t=\backslash frac\{\backslash Delta\_S-\backslash Delta\_T\}\{S\_t\}\; =\backslash frac\{v\_\backslash theta(z,t,c\_n)-v\_0(z,t,c\_p)\}\{S\_t\}.$$

Each $S_{t}S\_t$ is fitted only on training targets. First compute the per-coordinate sample standard deviation of ${\mathrm{\Delta }}_T\backslash Delta\_T$, floored at ${10}^{-4}10^\{-4\}$. Multiply that scale by the median per-row RMS of the normalized edits, so their median RMS becomes one. Development data never fits normalization.

The paired-error game

Use the same Gaussian noise vector for real and fake within a pair:

$$x_r={\sigma }_t(k)ϵ,x_f={\sigma }_t(k)ϵ+e_t,ϵ\sim \mathcal{N}(0,I)\mathrm{.}x\_r=\backslash sigma\_t(k)\backslash epsilon,\backslash qquad\; x\_f=\backslash sigma\_t(k)\backslash epsilon+e\_t,\backslash qquad\; \backslash epsilon\backslash sim\backslash mathcal\; N(0,I).$$

The critic is the pinned global-mix network: a projection of the full velocity residual into eight width-48 tokens, one attention layer with four heads, and output score $8\tanh (\cdot \mathrm{/}8)8\backslash tanh(\backslash cdot/8)$. Separate saved streams choose the D/G rows and noise.

$$L_D=\mathbb{E}\mathrm{softplus}(D(x_f)-D(x_r))+L_{\mathrm{c}\mathrm{a}\mathrm{p}},L_G=\mathbb{E}\mathrm{softplus}(D(x_r)-D(x_f))+L_{\mathrm{V}\mathrm{I}\mathrm{C}}\mathrm{.}L\_D=\backslash mathbb\; E\backslash ,\backslash operatorname\{softplus\}(D(x\_f)-D(x\_r))+L\_\{\backslash mathrm\{cap\}\},\backslash qquad\; L\_G=\backslash mathbb\; E\backslash ,\backslash operatorname\{softplus\}(D(x\_r)-D(x\_f))+L\_\{\backslash mathrm\{VIC\}\}.$$

Every fourth update the critic receives the exact autograd gradient cap, with lazy multiplier four:

$$L_{\mathrm{c}\mathrm{a}\mathrm{p}}=\frac{4}{2}\mathbb{E}[\max (0,\mathrm{\parallel }\mathrm{\nabla }D(x_r){\mathrm{\parallel }}_2-1{)}^2+\max (0,\mathrm{\parallel }\mathrm{\nabla }D(x_f){\mathrm{\parallel }}_2-1{)}^2]\mathrm{.}L\_\{\backslash mathrm\{cap\}\}=\backslash frac\{4\}\{2\}\backslash mathbb\; E\backslash left[\; \backslash max(0,\backslash |\backslash nabla\; D(x\_r)\backslash |\_2-1)^2+\; \backslash max(0,\backslash |\backslash nabla\; D(x\_f)\backslash |\_2-1)^2\backslash right].$$

Other updates have zero cap loss. VIC acts on a random 64-particle subset: mean $\max (0,1-\sqrt{\mathrm{Var}(P_j)+{10}^{-4}})\backslash max(0,1-\backslash sqrt\{\backslash operatorname\{Var\}(P\_j)+10^\{-4\}\})$ plus the squared off-diagonal sample covariance sum divided by particle dimension four. Both regularizer coefficients are one. There is no reconstruction or perceptual auxiliary loss in the particle training game.

Noise starts at the training edit RMS divided by 0.28. It decays geometrically toward 0.03 over 1,600 updates, with an absolute hold at 1.0 when the start exceeds that hold. Update $kk$ uses schedule index $k-1k-1$. The pilot keeps the same horizon.

Released recipe and evidence

Setting	Value
Checkpoint	EMA after 1,600 updates, seed 7
Adam learning rates	Generator $2\times {10}^{-5}2\backslash times10^\{-5\}$; critic $9\times {10}^{-4}9\backslash times10^\{-4\}$; particles $6\times {10}^{-3}6\backslash times10^\{-3\}$
Adam betas / decay	$(0,0.999)(0,0.999)$ / zero
Effective batch / microbatch	8 / 1
EMA decay	0.995
Training resolution	512 × 512
Public samples	768 × 768, 10 Euler steps, CFG 1, scheduler shift 3
Per slider training data	24 paired prompt rows × 16 seeds; both trajectories × 10 positions

The original adult character catalog separates 12 training characters, four development characters, and eight final-test characters. Six definitions train each slider; two extra paraphrases are held out. Candlelit and Moonlit edit lighting; Bad Intent uses expression, pose, framing and atmosphere definitions. The public gallery uses development cases, including a bare prompt, and is not a final-test benchmark.

These are final-budget experimental checkpoints. They were not selected as the best checkpoint by perceptual quality. Strengths above one extrapolate beyond training and can change clothing, pose, setting, or composition. The measured 1,600-update runs did not establish convergence. Training traces, normalization provenance, checkpoint hashes, and development measurements accompany the weights.

Shared algorithm implementation

The reusable routing, global-mix critic, paired losses, particle regularizer, noise function and ordinary-LoRA solver live in sliders-conceptmod. This release pins that revision in requirements.txt and core.lock.json. Anima owns the model integration, target collection, training recipe and sample evidence. The extracted reference implementation is byte-identical to the original; runtime identity hashes the implementation rather than its compatibility import. Architecture and research.

License

The adapters are derivatives of CircleStone Anima and are distributed under the CircleStone Labs Non-Commercial License. Rights are granted directly by CircleStone Labs LLC. See NOTICE. Independently authored source code is MIT; vendored reference code retains its included license. This is an independent ntc-ai release.