My starting images were a photograph of my own refrigerator and an AI generated, photographic style close up of a halved orange. I chose these images because their recognizable structures could also suggest other forms: refrigerator shelves could become occupied floors, while orange segments could become mechanical chambers.
In response to last week’s feedback, I focused on transformations that retained traces of the original objects while introducing unfamiliar spatial or material relationships. I used XY Plot comparisons to investigate denoise strength, CFG, sampling methods, and checkpoints.
Experiment 1: Transforming a Refrigerator into an Architectural Space
Original Image
Research question: How do denoise strength and CFG affect the transformation of a familiar household scene into an imagined architectural space?
My starting image was a photograph of my own refrigerator. I chose it because its shelves, containers, and enclosed interior provided recognizable structures that could be reinterpreted as floors, rooms, and circulation spaces.
I conducted three attempts, moving from a broad denoise range to a narrower comparison. Each generation started from the original photograph. I kept the model, seed, sampler, scheduler, and steps setting unchanged. After the first attempt, I revised the positive prompt and used that revised version for Attempts 2 and 3.
The basic workflow was:
Load Image + Load Checkpoint + two CLIP Text Encode (Prompt) nodes → VAE Encode and KSampler (Efficient), controlled by XY Input: Denoise + XY Input: CFG Scale → XY Plot → saved through KSampler’s IMAGE output → Save Image.
Data Information：
Model: sd_xl_base_1.0.safetensors
Seed: 12345
Control after Generate: fixed
Steps setting: 20
Sampler: dpm_adaptive
Scheduler: normal
Input image size: 1264 × 1670
X-axis: Denoise
The CFG and denoise values were supplied by the XY input nodes, overriding the values displayed directly in KSampler.
Attempt 1: Broad Exploration
Denoise: 0.35 / 0.60 / 0.85
CFG: 3.0 / 5.0 / 7.0
Y-axis: CFG
VAE decode: true
XY Plot output: Plot
Positive prompt:
“Photograph of a towering inhabited atrium. Three stacked floors connected by narrow staircases, walkways with railings, doorways opening into deep rooms. Two small people stand on separate levels. Walls made of translucent ribbed plastic, crumpled silver foil partitions, pale ceramic surfaces, cold overhead lighting. Frontal view, realistic materials, strong spatial depth.”
Negative prompt:
“blur,text,cartoon,watercolor”
Observations
0.35 and 0.60 — The refrigerator remains recognizable.
The results stayed close to the source photograph. Changes mainly affected packaging, containers, and local surface details.
0.85 — Reorganized storage.
The arrangement became more compartmentalized, with additional containers and divisions. However, the image still read as refrigerator storage rather than an inhabited architectural space.
The initial prompt did not produce the intended change of scale. I revised it to describe an atrium, staircases, and people more directly.
Attempt 2: Narrowing the Range
Denoise: 0.45 / 0.55 / 0.65
CFG: 3.0 / 5.0 / 7.0
Positive prompt:
“Photograph of a towering inhabited atrium. Three stacked floors connected by narrow staircases, walkways with railings, doorways opening into deep rooms. Two small people stand on separate levels. Walls made of translucent ribbed plastic, crumpled silver foil partitions, pale ceramic surfaces, cold overhead lighting. Frontal view, realistic materials, strong spatial depth.”
Negative prompt:
“blur,text,cartoon,watercolor”
Observations
0.45 — Household scale.
Food packages and containers remained identifiable, and the scene still appeared primarily as a refrigerator.
0.55 — Coexisting spatial interpretations.
Figures, staircases, and occupied floors appeared within the original framework. The refrigerator border and traces of packaging remained, producing an ambiguous relationship between a small storage space and a large interior.
0.65 — More developed architecture.
The architectural organization became clearer, while individual food objects became less recognizable. The original shelves continued to influence the arrangement of floors.
Since both the prompt and the denoise range changed after Attempt 1, I treated this as a new exploratory stage rather than attributing the difference to denoise alone.
Attempt 3: Examining the Transition
Denoise: 0.48 / 0.51 / 0.54
CFG: 3.0 / 5.0 / 7.0
Positive prompt:
“Photograph of a towering inhabited atrium. Three stacked floors connected by narrow staircases, walkways with railings, doorways opening into deep rooms. Two small people stand on separate levels. Walls made of translucent ribbed plastic, crumpled silver foil partitions, pale ceramic surfaces, cold overhead lighting. Frontal view, realistic materials, strong spatial depth.”
Negative prompt:
“blur,text,cartoon,watercolor”
Observations
0.48 — CFG influences the transition.
At CFG 3, the image remained close to a refrigerator. At CFG 5 and 7, architectural features appeared. This comparison showed that the transition depended on the combination of denoise and CFG.
0.51 — A container becomes a staircase.
The round food container developed into a circular structure with stairs. Its shape still recalled the original object, while the steps and small figures suggested an occupiable space.
0.54 — Continued architectural interpretation.
The results remained close to those at 0.51, with changes in figures, surfaces, and local structural details. Increasing denoise within this interval did not produce an entirely new composition.
Reflection
The most interesting result was the transformation of a food container into a circular staircase at denoise 0.51 and CFG 5. It preserved a visible relationship to the source photograph while changing the object’s apparent scale and function.
The three attempts helped me move from general variation toward a specific visual question: when does a familiar object begin to support a different interpretation? Denoise and CFG jointly affected this transition, while the prompt revision changed the direction of the experiment.
These findings are limited to this input photograph, model, workflow, and fixed seed. I did not test whether the same transitions would persist across other seeds.
Experiment 2: Comparing Samplers and Sampling Steps
Research question: How do different samplers and step settings affect image clarity and the preservation of the refrigerator’s transformation into an architectural space?
I used the same original refrigerator photograph and the revised atrium prompt from Experiment 1. I chose denoise 0.51 and CFG 5 because this combination produced a circular staircase that still resembled the original food container.
I compared three samplers at three step settings in a single nine-image XY Plot. The input image, positive and negative prompts, model, seed, scheduler, denoise, and CFG remained constant.
The basic workflow was:
Load Image + Load Checkpoint + two CLIP Text Encode (Prompt) nodes → VAE Encode and KSampler (Efficient), with XY Input: Sampler/Scheduler + XY Input: Steps → XY Plot feeding the sampler’s script input → KSampler’s IMAGE output → Save Image.
Data Information
Model: sd_xl_base_1.0.safetensors
Seed: 12345
Control after Generate: fixed
CFG: 5.0
Denoise: 0.51
Scheduler: normal
Input image size: 1264 × 1670
X-axis: Sampler — euler / dpmpp_2m / dpmpp_2m_sde
Y-axis: Steps — 20 / 30 / 40
VAE decode: true
XY Plot output: Plot
The sampler and steps values were supplied by the XY input nodes, overriding the corresponding settings displayed directly in KSampler.
Positive prompt
“Photograph of a towering inhabited atrium. Three stacked floors connected by narrow staircases, walkways with railings, doorways opening into deep rooms. Two small people stand on separate levels. Walls made of translucent ribbed plastic, crumpled silver foil partitions, pale ceramic surfaces, cold overhead lighting. Frontal view, realistic materials, strong spatial depth.”
Negative prompt
“blur,text,cartoon,watercolor”
Observations
Euler — Stable composition across step settings.
At 20, 30, and 40 steps, the overall arrangement remained similar. The circular staircase retained the outline of the original food container, and the refrigerator shelves continued to function as architectural floors. Increasing steps produced relatively subtle changes and did not noticeably eliminate the soft appearance of the interior.
DPM++ 2M — Similar structure and material continuity.
The results were visually close to those produced by Euler. The circular structure, upper-level figures, and lower passage remained in similar positions. Differences between 20, 30, and 40 steps were limited, and the grid did not show a clear, consistent improvement in sharpness as the step count increased.
DPM++ 2M SDE — More noticeable local reinterpretation.
This sampler produced more visible changes in the distribution of people, the right-hand wall, and the circular structure. At 20 steps, surfaces appeared mottled and fragmented. At 30 and 40 steps, they appeared more even, although some figures and railings still lacked complete, clearly defined forms.
The circular structure also appeared more like a manufactured architectural cylinder, retaining less of the food container’s original material character.
Reflection
Increasing steps from 20 to 40 did not consistently improve clarity in this experiment. Additional texture sometimes made the image more visually complicated without making its structures easier to understand.
The sampler comparison also revealed a trade-off between local reinterpretation and continuity with the source photograph. Euler and DPM++ 2M preserved more of the container-like appearance of the circular staircase, while DPM++ 2M SDE introduced more changes to the surrounding space.
For my purposes, a successful result needed both readable details and an interesting relationship to the original object. I therefore did not select a sampler solely on the basis of apparent sharpness. These observations are limited to the tested image, prompt, settings, and seed; they do not establish a general ranking of the three samplers.
Experiment 3: Comparing Models through an Organic-to Mechanical Transformation
Original Image
Research question: How do different checkpoints reinterpret the same organic structure when prompted to generate a mechanical object?
My starting image was an AI generated, photographic style close up of a halved orange. I chose it because its radial segments, white fibers, central core, and outer rind could suggest mechanical chambers, connections, a shaft, and a housing.
I compared two checkpoints at three denoise values. I conducted two attempts with different positive prompts, producing a six-image XY Plot for each attempt. Each generation started from the original orange image. The seed, sampler, steps, scheduler, CFG, and negative prompt remained unchanged.
The basic workflow was:
Load Image → VAE Encode → KSampler (Efficient), supplied by Efficient Loader and controlled by XY Input: Checkpoint + XY Input: Denoise → XY Plot → KSampler’s IMAGE output → Save Image.
Efficient Loader supplied the model, positive and negative conditioning, VAE, and XY Plot dependencies. I entered both prompts directly into Efficient Loader. Its empty LATENT output was not used; the sampler received the encoded orange image from VAE Encode.
Data Information
Checkpoints:
v1-5-pruned-emaonly.safetensors
realisticVisionV60B1_v51VAE.safetensors
Seed: 12345
Control after Generate: fixed
Steps: 30
CFG: 5.0
Sampler: dpmpp_2m
Scheduler: normal
VAE setting: Baked VAE
Clip skip: -1
LoRA: None
Input image size: 1254 × 1254
X-axis: Checkpoint
Y-axis: Denoise — 0.45 / 0.65 / 0.85
VAE decode: true
XY Plot output: Plot
The checkpoint and denoise values were supplied by the XY input nodes.
Attempt 1: Combining Organic Tissue and Mechanical Components
Positive prompt:
“Close-up photograph of a cut-open biomechanical device. Radial chambers of translucent amber tissue contain small pistons and interlocking mechanisms. White fibrous membranes connect the chambers to a central hub. A porous orange outer shell encloses the device. Organic tissue and mechanical parts form one continuous structure. Resting on a gray tabletop, side lighting, realistic material textures.”
Negative prompt:
“text, watermark, cartoon, illustration, blurry”
Observations
0.45 — The original fruit remains recognizable.
Both models preserved the orange’s overall appearance. Differences mainly involved the arrangement of white fibers and the texture of the fruit’s surface.
0.65 — Different interpretations of organic texture.
SD 1.5 produced a finer, irregular network across the cut surface. Realistic Vision emphasized branching white fibers and more distinct radial divisions. Neither result contained clearly developed mechanical components.
0.85 — Material changes without a clear mechanical transformation.
SD 1.5 produced a granular, gel-like surface with a less distinct segment structure. Realistic Vision retained a clearer radial organization, with a web-like center and a more regular outer edge.
This first attempt mainly transformed the orange’s texture. I revised the prompt to describe a specific mechanical object and recognizable components more directly.
Attempt 2: Describing a Cutaway Mechanical Pump
Denoise: 0.45 / 0.65 / 0.85
Checkpoints and other settings: unchanged from Attempt 1.
Positive prompt:
“Macro photograph of a cutaway mechanical pump on a gray tabletop. A circular housing encloses six radial piston cylinders, interlocking steel gears, small bolts, copper tubes and a central metal shaft. Translucent orange membranes stretch between the metal components, with moist fibers attached to the joints. Precise machined edges, realistic metal reflections, side lighting, three-quarter view.”
Negative prompt:
“text, watermark, cartoon, illustration, blurry”
Observations
0.45 — Organic identity retained.
Both checkpoints still produced recognizable orange halves. SD 1.5 remained close to the original fruit, while Realistic Vision introduced more visible changes to the white fibers and outer rind.
0.65 — Different degrees of structural transformation.
SD 1.5 placed a small metallic component at the center while leaving most of the surrounding fruit recognizable.
Realistic Vision generated a more developed central assembly. The fruit’s radial divisions became transparent compartments, and the rind began to resemble a manufactured circular housing.
0.85 — Two distinct material outcomes.
SD 1.5 produced an intricate combination of translucent orange material and mechanical parts. The result retained a moist, irregular quality that could suggest fruit, resin, or a machine’s internal structure.
Realistic Vision produced a more clearly defined metallic housing, with a central opening, radial supports, and small mechanical components. The orange color and circular composition remained, but much of the fruit’s material identity disappeared.
Reflection
The revised prompt produced a clearer mechanical transformation and made the differences between the checkpoints more visible. At denoise 0.65, Realistic Vision generated a developed mechanical assembly, while SD 1.5 mainly introduced a central metal component. At 0.85, SD 1.5 retained more ambiguous organic material, while Realistic Vision produced a more recognizable manufactured object.
My two most interesting results were Realistic Vision at 0.65 and SD 1.5 at 0.85. They represented different kinds of coexistence between organic and mechanical forms. The first offered a more readable structure, while the second was harder to classify. When evaluating model differences, however, I compared results at matching denoise values within each grid.
This experiment suggested that a more complete mechanical transformation is not necessarily the most experimental result. The uncertain boundary between fruit, translucent material, and machinery was particularly relevant to my response to last week’s feedback.
These observations apply to the tested input, prompts, resolution, seed, and loaded checkpoint configurations, including their associated VAE behavior. They do not isolate training differences or establish that one model is generally better.