LKFMixer: Dual Systems for LLM Mixing & Super-Resolution

• LKFMixer is a term describing two distinct systems: one an HCI instrument that uses audio mixer metaphors for LLM personality control, and the other a CNN model for image super-resolution using large kernels.
• The tangible system employs motorized faders, rotary knobs, and presets to offer embodied, direct manipulation of LLM characteristics, enhancing creative control.
• The CNN-based model leverages decomposed 31×31 convolutions with feature modulation and selection to efficiently approximate non-local interactions and boost reconstruction performance.

In current arXiv usage, LKFMixer refers to two distinct systems rather than a single research line: a tangible “personality mixer” for LLMs described as the Memetic Mixer in HCI and NIME-oriented work, and a pure CNN-based single-image super-resolution model family built around very large convolution kernels for efficient non-local feature modeling (McNamara et al., 16 Apr 2025, Tang et al., 15 Aug 2025). The first appropriates the hardware language of analogue synthesisers and audio mixing—faders, knobs, presets, filters, and effects—to control intangible aspects of an LLM’s behavior and persona. The second uses a $31 \times 31$ large-kernel design, coordinate decomposition, and feature modulation/selection blocks to approximate the non-local modeling power of self-attention while remaining lightweight and fast.

1. Terminological scope and disambiguation

A concise way to distinguish the two uses is to separate them by domain and operative meaning.

Usage	Domain	Core idea
Memetic Mixer / LKFMixer	HCI, NIME, LLM interfaces	Audio-mixer metaphors for controlling LLM personality and style
LKFMixer	Image restoration, super-resolution	Large-kernel CNN for efficient non-local SR

A plausible source of confusion is that both systems use the language of mixing, but in different senses. In the HCI system, mixing is literal at the level of interface metaphor: users “mix” personality traits, filters, and effects through physical controls. In the super-resolution model, mixing occurs inside the network through modules that fuse local and non-local image features. The shared name therefore does not indicate a shared architecture, task, or research group; it identifies two unrelated artifacts that should be cited separately by arXiv identifier when precision matters (McNamara et al., 16 Apr 2025, Tang et al., 15 Aug 2025).

2. LKFMixer as a tangible LLM “personality mixer”

In the HCI usage, LKFMixer is the Memetic Mixer, a bespoke tabletop device that borrows the visual and interaction language of a Moog-style analogue synthesiser and a small audio mixing console to control a LLM (McNamara et al., 16 Apr 2025). Its form factor comprises a tile recognition slate for physical word tiles, a set of five motorised faders, multiple rotary knobs grouped by function, and an e-Paper/e-ink screen that displays both the text assembled from the tiles and the LLM’s response.

The control vocabulary is explicitly audio-derived. Faders provide smooth, continuous control over personality traits; knobs are used for parameter tweaking and exploration of extremes; presets recall saved combinations; filters apply selective constraints; and effects “colour” the response. The authors frame the system as a “personality mixer” in which the user physically blends multiple personality traits of the model in a manner analogous to audio-track mixing.

The five motorised faders implement bipolar personality axes:

• Optimist — Pessimist
• Dreamer — Practical
• Dominant — Submissive
• Playful — Serious
• Trusting — Suspicious

These are represented conceptually as a vector

$$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$

with each scalar $x_k \in [-1,1]$. The paper does not publish the exact mapping code, but the device maps the current fader state into natural-language persona descriptions used in system prompts.

The device also exposes two preset layers. A Mode Presets knob selects one of four creative-process stages—Explore, Compose, Critique, and Refine—which switch the internal prompt-chain pattern. A separate personality preset control recalls archetypal personas: Guru, Bestie, Accountant, and Cynic. Selecting a preset moves the motorised faders to stored positions, mirroring the scene-recall behavior of motorised faders on digital consoles.

Four filter knobs apply additional constraints without changing the base persona or mode: Age, Vocabulary, Length, and Temperature. Three are encoded as prompt constraints, while Temperature directly manipulates the LLM’s sampling temperature: $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$ where $y_{\text{temp}} \in [0,1]$. Four effects knobs—Morality, Politics, Distress, and Sarcasm—act as tonal modifiers layered on top of the current personality mix.

This design is positioned against standard chatbot interfaces, which the authors describe as transactional, transitory and superficial, and poor at supporting slower, reflective creative interaction. The Memetic Mixer instead aims for a more special and purposeful interaction mode, grounded in tangible HCI, direct manipulation, and tinkerability in the sense associated with Resnick et al.

3. Interaction loop, qualitative evaluation, and HCI implications

The Memetic Mixer uses OpenAI’s ChatGPT-4o as the LLM backend via the OpenAI API, with no participant-specific identities sent (McNamara et al., 16 Apr 2025). The interaction loop begins with physical word tiles placed on the slate. The system detects the tiles and reconstructs a short poetic prompt text from a limited vocabulary, continuing the earlier Mimetic Poet design. That tile text is then routed through AI Chains-style prompt chaining, following Wu et al. and prior work by McCormack et al., to generate richer responses such as analogy, allegory, or ideation. Current fader positions, preset states, filter settings, effect settings, and generation parameters are converted into structured control information that modifies the chain. The resulting composite prompt is sent to ChatGPT-4o, and the textual output is rendered on the e-Paper screen. Users can then leave the tiles unchanged and iteratively manipulate the control surface to explore a space of alternative responses.

The comparative study sets the Memetic Mixer against the Mimetic Poet, a non-mixer device with word tiles and minimal markers but no knobs, faders, or audio metaphors. The study involved $N = 6$ Fine Art Honours students, aged 24–57, in a shared art studio. It was an ecological, performance-led / in-the-wild design: each device was installed for one week in the habituated studio environment, participants used them whenever and however they saw fit, and each week concluded with a 1-hour focus group comprising semi-structured interviews and open discussion. Data collection included audio recordings, transcripts, researcher notes, and still photography. Analysis used reflexive, inductive thematic analysis following Braun & Clarke. No quantitative measures such as Likert ratings or usage logs are reported.

The reported findings distinguish the two variants sharply. For the non-mixer device, participants appreciated the physicality but described rapid loss of engagement, limited expressive range, and outputs that often paraphrased the input too closely. They expressed a desire for more associative, cross-disciplinary, and unconventional responses, and they spontaneously proposed audio metaphors such as “knobs and sliders” to dial traits like “snarkiness” or “verbosity” up and down. For the Memetic Mixer, participants reported greater expressive range and playfulness, support for both fine-tuning and extreme parameter exploration, and a stronger incentive for continued interaction. They described the personality sliders as intuitive, welcomed the absence of instructions, and characterized the device as enabling a more malleable discussion with the LLM and the ability to shape it in front of you.

The paper’s central result is that audio-like controls afforded more immediate, direct and embodied control over the LLM than the non-mixer variant. The authors attribute this to visible and tactile state representation, direct manipulative mappings, immediate iterative feedback, cultural familiarity with mixer-like controls, and the avoidance of prompt-engineering burden. Within HCI and NIME, the system is situated alongside Ishii & Ullmer’s tangible bits, Ullmer & Ishii’s frameworks, metaphor-as-design-tool work by Blackwell and Bau et al., and entanglement-oriented perspectives connecting HCI and musical practice.

The limitations are explicit. The tile-only input vocabulary both supports creativity and restricts expressiveness. The sample is small and purely qualitative. The mappings from physical controls to prompt-engineering strategies are hand-designed and may require updating as LLMs evolve. Accessibility for blind and low-vision users remains an open issue. The authors also note that broader populations may not respond to the mixer metaphor as the studio cohort did.

4. LKFMixer as a large-kernel CNN for efficient image super-resolution

In the vision usage, LKFMixer is a family of pure CNN-based single-image super-resolution models that use very large convolution kernels $(31 \times 31)$ to simulate the non-local feature modeling of Transformer self-attention while keeping the efficiency of lightweight CNNs (Tang et al., 15 Aug 2025). The central problem is the familiar tension in SR: good reconstruction requires both strong local feature extraction and access to non-local information, but self-attention is computationally heavy and memory-hungry. The model therefore asks whether a CNN can recover Transformer-like non-local capacity through large kernels rather than attention.

The architecture follows a standard SR pipeline. A shallow $3 \times 3$ convolution maps the low-resolution input $I_{LR}$ to shallow feature $F_0$. Deep processing then proceeds through a sequence of Feature Modulation Blocks (FMBs). Each FMB contains a Feature Distillation Block (FDB), a Spatial Feature Modulation Block (SFMB), and a Feature Selection Block (FSB). Reconstruction is performed with another $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$0 convolution followed by sub-pixel convolution (PixelShuffle) for $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$1, $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$2, or $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$3 upsampling.

The family includes three variants:

• LKFMixer-T: 40 channels and 6 FMBs
• LKFMixer-B: 48 channels and 8 FMBs
• LKFMixer-L: 64 channels and 12 FMBs

All variants share the same large-kernel size 31 and the same channel split factor $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$4.

The architectural core is the Partial Large Kernel Block (PLKB). A naïve $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$5 depthwise convolution is expensive, so the model combines three reductions: depthwise convolution, operation on only a fraction of channels, and coordinate decomposition into stripe convolutions: $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$6 Given an input feature map $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$7, the first $$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$8 channels are processed with the decomposed large-kernel operation,

$$\mathbf{p} = (x_{\text{opt-pess}}, x_{\text{dream-prac}}, x_{\text{dom-sub}}, x_{\text{play-ser}}, x_{\text{trust-susp}})$$9

the remaining channels are passed through unchanged, and the result is concatenated and fused with a $x_k \in [-1,1]$0 convolution. This reduces the complexity from $x_k \in [-1,1]$1 to approximately $x_k \in [-1,1]$2 per channel direction and keeps the heavy operation confined to 25\% of channels.

Three higher-level modules organize feature interaction. The Feature Fusion Block (FFB) combines local features from DWConv$x_k \in [-1,1]$3 with non-local features from PLKB. The Feature Distillation Block (FDB) applies a multi-stage distillation structure in which three sequential FFBs form a refine branch and three $x_k \in [-1,1]$4 convolutions form a distillation branch; distilled features and the last refine output are then concatenated and fused. The Spatial Feature Modulation Block (SFMB) combines coarse spatial abstraction via Adaptive-Max-Pooling and bilinear upsampling with channel weighting from Adaptive-Avg-Pooling and sigmoid activation, then fuses this with PLKB output. The Feature Selection Block (FSB) performs complementary gating between local and non-local branches: $x_k \in [-1,1]$5 with $x_k \in [-1,1]$6 produced by a sigmoid-activated $x_k \in [-1,1]$7 convolution over concatenated branch features.

The model’s relation to self-attention is therefore functional rather than algorithmic. Large kernels supply a wide receptive field, while SFMB and FSB modulate and balance local and non-local information. The paper explicitly treats this as a pure-convolution alternative to attention-heavy lightweight SR.

5. Training regime, empirical performance, and efficiency

The training configuration uses DIV2K and Flickr2K for training, with low-resolution images generated by bicubic downsampling, and evaluates on Set5, Set14, BSD100, Urban100, and Manga109 (Tang et al., 15 Aug 2025). Training is patch-based: high-resolution images are cut into $x_k \in [-1,1]$8 patches via sliding windows, with $x_k \in [-1,1]$9 low-resolution patches in the training pipeline, and random horizontal/vertical flips and rotations are used for augmentation. The loss combines $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$0 loss and FFT loss,

$$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$1

and optimization uses Adam for 1,000K iterations with an initial learning rate of $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$2, a minimum of $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$3, and cosine annealing. Evaluation is performed on the Y-channel in YCbCr.

For $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$4 SR, the three model sizes are:

• LKFMixer-T: 207K params, 11G FLOPs
• LKFMixer-B: 373K params, 19.9G FLOPs
• LKFMixer-L: 927K params, 49.5G FLOPs

Against SwinIR-light at $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$5, which has 930K parameters and 65G FLOPs, LKFMixer-L retains a similar parameter count with lower FLOPs. On an RTX 3090 with $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$6 input at $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$7, the runtime and memory figures are: 405 MB / 61 ms for LKFMixer-T, 486 MB / 99 ms for LKFMixer-B, and 648 MB / 194 ms for LKFMixer-L. By comparison, SwinIR-light uses 1451 MB and 1000 ms, SRFormer-light uses 1304 MB and 998 ms, CAMixerSR-light uses 1489 MB and 312 ms, and SeeMoRe-L uses 471 MB and 233 ms.

The headline quantitative result is the $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$8 comparison between LKFMixer-L and SwinIR-light. LKFMixer-L reports:

• Set5: 32.71 / 0.9008
• Set14: 28.95 / 0.7892
• BSD100: 27.83 / 0.7443
• Urban100: 26.85 / 0.8069
• Manga109: 31.52 / 0.9191

Against SwinIR-light, the gains are +0.27 dB on Set5, +0.18 dB on Set14, +0.14 dB on BSD100, +0.38 dB on Urban100, and +0.60 dB on Manga109, while inference is about $$T = T_{\min} + y_{\text{temp}} \cdot (T_{\max} - T_{\min})$$9 faster. At $y_{\text{temp}} \in [0,1]$0, the paper reports 34.63 / 0.9503 on Manga109 for LKFMixer-L, ahead of SwinIR-light, MAN-light, and SeeMoRe-L. The authors also report competitive or better LPIPS values, including 0.1001 on Manga109 versus 0.1021 for SwinIR-light.

The empirical analysis includes Local Attribution Map (LAM) and Effective Receptive Field (ERF) visualization. These show that LKFMixer-L attains higher Diffusion Index and a larger effective receptive field than SwinIR-light and other lightweight SR models. Ablation results support the design. Replacing the decomposed PLKB with full $y_{\text{temp}} \in [0,1]$1 depthwise convolution increases parameters by +115.5\% and inference time by +97.7\% with no meaningful performance gain. Replacing PLKB with window-based self-attention produces only marginal PSNR improvement but raises parameters from 373K to 678K, memory from about 486M to about 1490M, and inference time from 98 ms to 1354 ms. Kernel-size ablation shows improvement up to 31, with 41 producing a decline. Removing FDB, SFMB, or FSB consistently reduces performance, and removing the PLKB branch from SFMB is especially damaging.

6. Limits, future directions, and broader significance

The two LKFMixer systems occupy different research traditions and point toward different open problems, but both are explicitly concerned with making high-dimensional behavior more tractable (McNamara et al., 16 Apr 2025, Tang et al., 15 Aug 2025). In the HCI system, the issue is how to give users direct, embodied access to intangible LLM properties such as persona, tone, and randomness. In the vision system, the issue is how to give a pure CNN access to broad non-local context without paying the runtime and memory costs associated with self-attention. This suggests that the shared name corresponds less to a common method than to a common design intuition: structured control through mixing.

For the Memetic Mixer, future work is framed around goal-oriented interactions, domain-specific differences in the appeal of audio metaphors, and accessibility. The study does not quantify productivity or creative output, and its ecological design prioritizes experience and engagement over controlled performance metrics. For the super-resolution model, the authors identify open questions around more sophisticated fusion of local and non-local features, the unexplained performance decline beyond kernel size 31, and extension to video SR, denoising, deblurring, and real-world SR. Its main experiments are centered on bicubic downsampling, while real-world degradation results remain qualitative.

Taken together, the two usages of LKFMixer illustrate a notable divergence in contemporary arXiv research vocabulary. One denotes a physical interface that repurposes audio-console metaphors for LLM interaction; the other denotes a large-kernel convolutional framework for efficient super-resolution. Precision therefore depends on explicit citation. In HCI and creative-AI contexts, LKFMixer names a tangible LLM instrument centered on faders, knobs, presets, filters, and effects. In low-level vision, LKFMixer names a lightweight SR family whose principal contribution is the efficient use of decomposed $y_{\text{temp}} \in [0,1]$2 large kernels with feature distillation, modulation, and selection.